program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func main<ios18>(tensor<int32, [1]> audio_length, tensor<fp32, [1, 128, 32]> audio_signal, tensor<fp32, [24, 1, 56, 1024]> cache_last_channel, tensor<int32, [1]> cache_last_channel_len, tensor<fp32, [24, 1, 1024, 8]> cache_last_time, tensor<fp32, [1, 128]> language_mask, tensor<fp32, [1, 128, 9]> pre_cache) {
            int32 var_10 = const()[name = string("op_10"), val = int32(2)];
            int32 var_18 = const()[name = string("op_18"), val = int32(-1)];
            bool full_input_1_interleave_0 = const()[name = string("full_input_1_interleave_0"), val = bool(false)];
            string pre_cache_to_fp16_dtype_0 = const()[name = string("pre_cache_to_fp16_dtype_0"), val = string("fp16")];
            string audio_signal_to_fp16_dtype_0 = const()[name = string("audio_signal_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 9]> pre_cache_to_fp16 = cast(dtype = pre_cache_to_fp16_dtype_0, x = pre_cache)[name = string("cast_17")];
            tensor<fp16, [1, 128, 32]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = string("cast_18")];
            tensor<fp16, [1, 128, 41]> full_input_1_cast_fp16 = concat(axis = var_10, interleave = full_input_1_interleave_0, values = (pre_cache_to_fp16, audio_signal_to_fp16))[name = string("full_input_1_cast_fp16")];
            int32 var_23 = const()[name = string("op_23"), val = int32(9)];
            tensor<int32, [1]> value_3 = add(x = audio_length, y = var_23)[name = string("value_3")];
            tensor<int32, [3]> var_27_begin_0 = const()[name = string("op_27_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> var_27_end_0 = const()[name = string("op_27_end_0"), val = tensor<int32, [3]>([1, 128, 41])];
            tensor<bool, [3]> var_27_end_mask_0 = const()[name = string("op_27_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 9]> new_pre_cache = slice_by_index(begin = var_27_begin_0, end = var_27_end_0, end_mask = var_27_end_mask_0, x = full_input_1_cast_fp16)[name = string("op_27_cast_fp16")];
            int32 var_56 = const()[name = string("op_56"), val = int32(-1)];
            int32 var_64 = const()[name = string("op_64"), val = int32(1)];
            tensor<int32, [3]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> tensor_2_axes_0 = const()[name = string("tensor_2_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 41, 128]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = full_input_1_cast_fp16)[name = string("transpose_362")];
            tensor<fp16, [1, 1, 41, 128]> tensor_2_cast_fp16 = expand_dims(axes = tensor_2_axes_0, x = x_3_cast_fp16)[name = string("tensor_2_cast_fp16")];
            tensor<int32, [1, 41]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1, 41]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<int32, [1]> var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_139 = expand_dims(axes = var_139_axes_0, x = value_3)[name = string("op_139")];
            tensor<bool, [1, 41]> time_mask_1 = less(x = expand_dims_0, y = var_139)[name = string("time_mask_1")];
            tensor<int32, [1]> var_141_axes_0 = const()[name = string("op_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 41, 1]> var_141 = expand_dims(axes = var_141_axes_0, x = time_mask_1)[name = string("op_141")];
            tensor<int32, [3]> var_143_reps_0 = const()[name = string("op_143_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
            tensor<bool, [1, 41, 128]> var_143 = tile(reps = var_143_reps_0, x = var_141)[name = string("op_143")];
            tensor<int32, [1]> var_149_axes_0 = const()[name = string("op_149_axes_0"), val = tensor<int32, [1]>([1])];
            string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 41, 128]> var_143_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = var_143)[name = string("cast_16")];
            tensor<fp16, [1, 1, 41, 128]> var_149_cast_fp16 = expand_dims(axes = var_149_axes_0, x = var_143_to_fp16)[name = string("op_149_cast_fp16")];
            tensor<fp16, [1, 1, 41, 128]> input_3_cast_fp16 = mul(x = tensor_2_cast_fp16, y = var_149_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<int32, [8]> input0_5_pad_0 = const()[name = string("input0_5_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
            string input0_5_mode_0 = const()[name = string("input0_5_mode_0"), val = string("constant")];
            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 44, 131]> input0_5_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input0_5_mode_0, pad = input0_5_pad_0, x = input_3_cast_fp16)[name = string("input0_5_cast_fp16")];
            string tensor_4_pad_type_0 = const()[name = string("tensor_4_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> tensor_4_strides_0 = const()[name = string("tensor_4_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [4]> tensor_4_pad_0 = const()[name = string("tensor_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_4_dilations_0 = const()[name = string("tensor_4_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 tensor_4_groups_0 = const()[name = string("tensor_4_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2688))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_palettized")];
            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3264)))];
            tensor<fp16, [1, 256, 21, 65]> tensor_4_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_4_dilations_0, groups = tensor_4_groups_0, pad = tensor_4_pad_0, pad_type = tensor_4_pad_type_0, strides = tensor_4_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_palettized, x = input0_5_cast_fp16)[name = string("tensor_4_cast_fp16")];
            string cast_0_to_fp16_dtype_0 = const()[name = string("cast_0_to_fp16_dtype_0"), val = string("fp16")];
            fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> value_3_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = value_3)[name = string("cast_15")];
            tensor<fp16, [1]> var_163_cast_fp16 = add(x = value_3_to_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")];
            fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> var_165_cast_fp16 = add(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")];
            fp16 var_166_promoted_to_fp16 = const()[name = string("op_166_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_167_cast_fp16 = sub(x = var_165_cast_fp16, y = var_166_promoted_to_fp16)[name = string("op_167_cast_fp16")];
            fp16 var_67_promoted_to_fp16 = const()[name = string("op_67_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_167_cast_fp16, y = var_67_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")];
            fp16 var_169_promoted_to_fp16 = const()[name = string("op_169_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> current_lengths0_1_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_169_promoted_to_fp16)[name = string("current_lengths0_1_cast_fp16")];
            string cast_3_dtype_0 = const()[name = string("cast_3_dtype_0"), val = string("int32")];
            tensor<int32, [1, 21]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1, 21]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3840)))];
            tensor<int32, [1]> var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> current_lengths0_1_cast_fp16_to_int32 = cast(dtype = cast_3_dtype_0, x = current_lengths0_1_cast_fp16)[name = string("cast_14")];
            tensor<int32, [1, 1]> var_178 = expand_dims(axes = var_178_axes_0, x = current_lengths0_1_cast_fp16_to_int32)[name = string("op_178")];
            tensor<bool, [1, 21]> time_mask0_1 = less(x = expand_dims_1, y = var_178)[name = string("time_mask0_1")];
            tensor<int32, [1]> var_180_axes_0 = const()[name = string("op_180_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 21, 1]> var_180 = expand_dims(axes = var_180_axes_0, x = time_mask0_1)[name = string("op_180")];
            tensor<int32, [3]> var_182_reps_0 = const()[name = string("op_182_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
            tensor<bool, [1, 21, 65]> var_182 = tile(reps = var_182_reps_0, x = var_180)[name = string("op_182")];
            tensor<int32, [1]> var_188_axes_0 = const()[name = string("op_188_axes_0"), val = tensor<int32, [1]>([1])];
            string cast_4_to_fp16_dtype_0 = const()[name = string("cast_4_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 21, 65]> var_182_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_182)[name = string("cast_13")];
            tensor<fp16, [1, 1, 21, 65]> var_188_cast_fp16 = expand_dims(axes = var_188_axes_0, x = var_182_to_fp16)[name = string("op_188_cast_fp16")];
            tensor<int32, [4]> expanded_mask0_1_reps_0 = const()[name = string("expanded_mask0_1_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
            tensor<fp16, [1, 256, 21, 65]> expanded_mask0_1_cast_fp16 = tile(reps = expanded_mask0_1_reps_0, x = var_188_cast_fp16)[name = string("expanded_mask0_1_cast_fp16")];
            tensor<fp16, [1, 256, 21, 65]> input0_7_cast_fp16 = mul(x = tensor_4_cast_fp16, y = expanded_mask0_1_cast_fp16)[name = string("input0_7_cast_fp16")];
            tensor<fp16, [1, 256, 21, 65]> var_192_cast_fp16 = relu(x = input0_7_cast_fp16)[name = string("op_192_cast_fp16")];
            tensor<fp16, [1, 256, 21, 65]> input1_6_cast_fp16 = mul(x = var_192_cast_fp16, y = expanded_mask0_1_cast_fp16)[name = string("input1_6_cast_fp16")];
            tensor<int32, [8]> input0_9_pad_0 = const()[name = string("input0_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
            string input0_9_mode_0 = const()[name = string("input0_9_mode_0"), val = string("constant")];
            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 24, 68]> input0_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input0_9_mode_0, pad = input0_9_pad_0, x = input1_6_cast_fp16)[name = string("input0_9_cast_fp16")];
            string tensor_6_pad_type_0 = const()[name = string("tensor_6_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> tensor_6_strides_0 = const()[name = string("tensor_6_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 tensor_6_groups_0 = const()[name = string("tensor_6_groups_0"), val = int32(256)];
            tensor<int32, [4]> tensor_6_pad_0 = const()[name = string("tensor_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_6_dilations_0 = const()[name = string("tensor_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6400))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_palettized")];
            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6976)))];
            tensor<fp16, [1, 256, 11, 33]> tensor_6_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_6_dilations_0, groups = tensor_6_groups_0, pad = tensor_6_pad_0, pad_type = tensor_6_pad_type_0, strides = tensor_6_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_palettized, x = input0_9_cast_fp16)[name = string("tensor_6_cast_fp16")];
            fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_211_cast_fp16 = add(x = current_lengths0_1_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")];
            fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> var_213_cast_fp16 = add(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")];
            fp16 var_214_promoted_to_fp16 = const()[name = string("op_214_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_215_cast_fp16 = sub(x = var_213_cast_fp16, y = var_214_promoted_to_fp16)[name = string("op_215_cast_fp16")];
            fp16 var_67_promoted_1_to_fp16 = const()[name = string("op_67_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_215_cast_fp16, y = var_67_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")];
            fp16 var_217_promoted_to_fp16 = const()[name = string("op_217_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> current_lengths1_1_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_217_promoted_to_fp16)[name = string("current_lengths1_1_cast_fp16")];
            string cast_5_dtype_0 = const()[name = string("cast_5_dtype_0"), val = string("int32")];
            tensor<int32, [1, 11]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1, 11]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7552)))];
            tensor<int32, [1]> var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> current_lengths1_1_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths1_1_cast_fp16)[name = string("cast_12")];
            tensor<int32, [1, 1]> var_226 = expand_dims(axes = var_226_axes_0, x = current_lengths1_1_cast_fp16_to_int32)[name = string("op_226")];
            tensor<bool, [1, 11]> time_mask1_1 = less(x = expand_dims_2, y = var_226)[name = string("time_mask1_1")];
            tensor<int32, [1]> var_228_axes_0 = const()[name = string("op_228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 11, 1]> var_228 = expand_dims(axes = var_228_axes_0, x = time_mask1_1)[name = string("op_228")];
            tensor<int32, [3]> var_230_reps_0 = const()[name = string("op_230_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
            tensor<bool, [1, 11, 33]> var_230 = tile(reps = var_230_reps_0, x = var_228)[name = string("op_230")];
            tensor<int32, [1]> var_236_axes_0 = const()[name = string("op_236_axes_0"), val = tensor<int32, [1]>([1])];
            string cast_6_to_fp16_dtype_0 = const()[name = string("cast_6_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 11, 33]> var_230_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_230)[name = string("cast_11")];
            tensor<fp16, [1, 1, 11, 33]> var_236_cast_fp16 = expand_dims(axes = var_236_axes_0, x = var_230_to_fp16)[name = string("op_236_cast_fp16")];
            tensor<int32, [4]> expanded_mask2_1_reps_0 = const()[name = string("expanded_mask2_1_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
            tensor<fp16, [1, 256, 11, 33]> expanded_mask2_1_cast_fp16 = tile(reps = expanded_mask2_1_reps_0, x = var_236_cast_fp16)[name = string("expanded_mask2_1_cast_fp16")];
            tensor<fp16, [1, 256, 11, 33]> input2_6_cast_fp16 = mul(x = tensor_6_cast_fp16, y = expanded_mask2_1_cast_fp16)[name = string("input2_6_cast_fp16")];
            string tensor_8_pad_type_0 = const()[name = string("tensor_8_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> tensor_8_strides_0 = const()[name = string("tensor_8_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> tensor_8_pad_0 = const()[name = string("tensor_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_8_dilations_0 = const()[name = string("tensor_8_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 tensor_8_groups_0 = const()[name = string("tensor_8_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73280))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_palettized")];
            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73856)))];
            tensor<fp16, [1, 256, 11, 33]> tensor_8_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_8_dilations_0, groups = tensor_8_groups_0, pad = tensor_8_pad_0, pad_type = tensor_8_pad_type_0, strides = tensor_8_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_palettized, x = input2_6_cast_fp16)[name = string("tensor_8_cast_fp16")];
            tensor<fp16, [1, 256, 11, 33]> input3_2_cast_fp16 = mul(x = tensor_8_cast_fp16, y = expanded_mask2_1_cast_fp16)[name = string("input3_2_cast_fp16")];
            tensor<fp16, [1, 256, 11, 33]> var_255_cast_fp16 = relu(x = input3_2_cast_fp16)[name = string("op_255_cast_fp16")];
            tensor<fp16, [1, 256, 11, 33]> input4_2_cast_fp16 = mul(x = var_255_cast_fp16, y = expanded_mask2_1_cast_fp16)[name = string("input4_2_cast_fp16")];
            tensor<int32, [8]> input0_11_pad_0 = const()[name = string("input0_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
            string input0_11_mode_0 = const()[name = string("input0_11_mode_0"), val = string("constant")];
            fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 256, 14, 36]> input0_11_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input0_11_mode_0, pad = input0_11_pad_0, x = input4_2_cast_fp16)[name = string("input0_11_cast_fp16")];
            string tensor_10_pad_type_0 = const()[name = string("tensor_10_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> tensor_10_strides_0 = const()[name = string("tensor_10_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 tensor_10_groups_0 = const()[name = string("tensor_10_groups_0"), val = int32(256)];
            tensor<int32, [4]> tensor_10_pad_0 = const()[name = string("tensor_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_10_dilations_0 = const()[name = string("tensor_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76800))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_palettized")];
            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77376)))];
            tensor<fp16, [1, 256, 6, 17]> tensor_10_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_10_dilations_0, groups = tensor_10_groups_0, pad = tensor_10_pad_0, pad_type = tensor_10_pad_type_0, strides = tensor_10_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_palettized, x = input0_11_cast_fp16)[name = string("tensor_10_cast_fp16")];
            fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_274_cast_fp16 = add(x = current_lengths1_1_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")];
            fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> var_276_cast_fp16 = add(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")];
            fp16 var_277_promoted_to_fp16 = const()[name = string("op_277_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_278_cast_fp16 = sub(x = var_276_cast_fp16, y = var_277_promoted_to_fp16)[name = string("op_278_cast_fp16")];
            fp16 var_67_promoted_2_to_fp16 = const()[name = string("op_67_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_278_cast_fp16, y = var_67_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")];
            fp16 var_280_promoted_to_fp16 = const()[name = string("op_280_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> current_lengths2_1_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_280_promoted_to_fp16)[name = string("current_lengths2_1_cast_fp16")];
            string cast_7_dtype_0 = const()[name = string("cast_7_dtype_0"), val = string("int32")];
            tensor<int32, [1, 6]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1, 6]>([[0, 1, 2, 3, 4, 5]])];
            tensor<int32, [1]> var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> current_lengths2_1_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths2_1_cast_fp16)[name = string("cast_10")];
            tensor<int32, [1, 1]> var_289 = expand_dims(axes = var_289_axes_0, x = current_lengths2_1_cast_fp16_to_int32)[name = string("op_289")];
            tensor<bool, [1, 6]> time_mask2_1 = less(x = expand_dims_3, y = var_289)[name = string("time_mask2_1")];
            tensor<int32, [1]> var_291_axes_0 = const()[name = string("op_291_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 6, 1]> var_291 = expand_dims(axes = var_291_axes_0, x = time_mask2_1)[name = string("op_291")];
            tensor<int32, [3]> var_293_reps_0 = const()[name = string("op_293_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
            tensor<bool, [1, 6, 17]> var_293 = tile(reps = var_293_reps_0, x = var_291)[name = string("op_293")];
            tensor<int32, [1]> var_299_axes_0 = const()[name = string("op_299_axes_0"), val = tensor<int32, [1]>([1])];
            string cast_8_to_fp16_dtype_0 = const()[name = string("cast_8_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 6, 17]> var_293_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_293)[name = string("cast_9")];
            tensor<fp16, [1, 1, 6, 17]> var_299_cast_fp16 = expand_dims(axes = var_299_axes_0, x = var_293_to_fp16)[name = string("op_299_cast_fp16")];
            tensor<int32, [4]> expanded_mask5_1_reps_0 = const()[name = string("expanded_mask5_1_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
            tensor<fp16, [1, 256, 6, 17]> expanded_mask5_1_cast_fp16 = tile(reps = expanded_mask5_1_reps_0, x = var_299_cast_fp16)[name = string("expanded_mask5_1_cast_fp16")];
            tensor<fp16, [1, 256, 6, 17]> input5_2_cast_fp16 = mul(x = tensor_10_cast_fp16, y = expanded_mask5_1_cast_fp16)[name = string("input5_2_cast_fp16")];
            string tensor_1_pad_type_0 = const()[name = string("tensor_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> tensor_1_strides_0 = const()[name = string("tensor_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> tensor_1_pad_0 = const()[name = string("tensor_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> tensor_1_dilations_0 = const()[name = string("tensor_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 tensor_1_groups_0 = const()[name = string("tensor_1_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143552))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_palettized")];
            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144128)))];
            tensor<fp16, [1, 256, 6, 17]> tensor_1_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_1_dilations_0, groups = tensor_1_groups_0, pad = tensor_1_pad_0, pad_type = tensor_1_pad_type_0, strides = tensor_1_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_palettized, x = input5_2_cast_fp16)[name = string("tensor_1_cast_fp16")];
            tensor<fp16, [1, 256, 6, 17]> input6_2_cast_fp16 = mul(x = tensor_1_cast_fp16, y = expanded_mask5_1_cast_fp16)[name = string("input6_2_cast_fp16")];
            tensor<fp16, [1, 256, 6, 17]> var_318_cast_fp16 = relu(x = input6_2_cast_fp16)[name = string("op_318_cast_fp16")];
            tensor<fp16, [1, 256, 6, 17]> x0_2_cast_fp16 = mul(x = var_318_cast_fp16, y = expanded_mask5_1_cast_fp16)[name = string("x0_2_cast_fp16")];
            tensor<int32, [4]> var_333_perm_0 = const()[name = string("op_333_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_334 = const()[name = string("op_334"), val = tensor<int32, [3]>([1, 6, -1])];
            tensor<fp16, [1, 6, 256, 17]> var_333_cast_fp16 = transpose(perm = var_333_perm_0, x = x0_2_cast_fp16)[name = string("transpose_361")];
            tensor<fp16, [1, 6, 4352]> input_5_cast_fp16 = reshape(shape = var_334, x = var_333_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4352]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144704))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601216))))[name = string("encoder_pre_encode_out_weight_to_fp16_palettized")];
            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601792)))];
            tensor<fp16, [1, 6, 1024]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> var_344_begin_0 = const()[name = string("op_344_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_344_end_0 = const()[name = string("op_344_end_0"), val = tensor<int32, [3]>([1, 6, 1024])];
            tensor<bool, [3]> var_344_end_mask_0 = const()[name = string("op_344_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 1024]> var_344_cast_fp16 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = linear_0_cast_fp16)[name = string("op_344_cast_fp16")];
            int32 var_346 = const()[name = string("op_346"), val = int32(2)];
            tensor<int32, [1]> var_347 = sub(x = current_lengths2_1_cast_fp16_to_int32, y = var_346)[name = string("op_347")];
            string var_347_promoted_to_fp16_dtype_0 = const()[name = string("op_347_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 var_61_promoted_to_fp16 = const()[name = string("op_61_promoted_to_fp16"), val = fp16(0x0p+0)];
            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1]> var_347_to_fp16 = cast(dtype = var_347_promoted_to_fp16_dtype_0, x = var_347)[name = string("cast_8")];
            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_61_promoted_to_fp16, beta = const_61_to_fp16, x = var_347_to_fp16)[name = string("clip_0_cast_fp16")];
            tensor<int32, [1]> max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor<int32, [1]>([4])];
            fp16 var_363_promoted_to_fp16 = const()[name = string("op_363_promoted_to_fp16"), val = fp16(0x1.cp+5)];
            tensor<fp16, [1]> padding_length_1_cast_fp16 = add(x = clip_0_cast_fp16, y = var_363_promoted_to_fp16)[name = string("padding_length_1_cast_fp16")];
            int32 const_63 = const()[name = string("const_63"), val = int32(-1)];
            tensor<int32, [1]> var_365 = mul(x = cache_last_channel_len, y = const_63)[name = string("op_365")];
            int32 var_366 = const()[name = string("op_366"), val = int32(56)];
            tensor<int32, [1]> offset_1 = add(x = var_365, y = var_366)[name = string("offset_1")];
            tensor<int32, [1]> var_406_axes_0 = const()[name = string("op_406_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> var_406_cast_fp16 = expand_dims(axes = var_406_axes_0, x = padding_length_1_cast_fp16)[name = string("op_406_cast_fp16")];
            tensor<fp16, [1, 60]> var_405_promoted_to_fp16 = const()[name = string("op_405_promoted_to_fp16"), val = tensor<fp16, [1, 60]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603904)))];
            tensor<bool, [1, 60]> pad_mask_1_cast_fp16 = less(x = var_405_promoted_to_fp16, y = var_406_cast_fp16)[name = string("pad_mask_1_cast_fp16")];
            tensor<int32, [1, 60]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1, 60]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604096)))];
            tensor<int32, [1]> var_412_axes_0 = const()[name = string("op_412_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 1]> var_412 = expand_dims(axes = var_412_axes_0, x = offset_1)[name = string("op_412")];
            tensor<bool, [1, 60]> pad_mask_off_1 = greater_equal(x = expand_dims_5, y = var_412)[name = string("pad_mask_off_1")];
            tensor<bool, [1, 60]> pad_mask0_1 = logical_and(x = pad_mask_off_1, y = pad_mask_1_cast_fp16)[name = string("pad_mask0_1")];
            tensor<int32, [1]> var_415_axes_0 = const()[name = string("op_415_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 60]> var_415 = expand_dims(axes = var_415_axes_0, x = pad_mask0_1)[name = string("op_415")];
            tensor<int32, [3]> var_416 = const()[name = string("op_416"), val = tensor<int32, [3]>([1, 60, 1])];
            tensor<bool, [1, 60, 60]> pad_mask_for_att_mask_1 = tile(reps = var_416, x = var_415)[name = string("pad_mask_for_att_mask_1")];
            tensor<int32, [3]> var_418_perm_0 = const()[name = string("op_418_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<bool, [1, 60, 60]> var_418 = transpose(perm = var_418_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_360")];
            tensor<bool, [1, 60, 60]> pad_mask_for_att_mask0_1 = logical_and(x = pad_mask_for_att_mask_1, y = var_418)[name = string("pad_mask_for_att_mask0_1")];
            tensor<bool, [1, 60, 60]> const_71 = const()[name = string("const_71"), val = tensor<bool, [1, 60, 60]>([[[true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
            tensor<bool, [1, 60, 60]> att_mask3_1 = logical_and(x = pad_mask_for_att_mask0_1, y = const_71)[name = string("att_mask3_1")];
            tensor<bool, [1, 60, 60]> att_mask4_1 = logical_not(x = att_mask3_1)[name = string("att_mask4_1")];
            tensor<bool, [1, 60]> pad_mask1_1 = logical_not(x = pad_mask0_1)[name = string("pad_mask1_1")];
            tensor<int32, [2]> pad_mask2_1_begin_0 = const()[name = string("pad_mask2_1_begin_0"), val = tensor<int32, [2]>([0, 56])];
            tensor<int32, [2]> pad_mask2_1_end_0 = const()[name = string("pad_mask2_1_end_0"), val = tensor<int32, [2]>([1, 60])];
            tensor<bool, [2]> pad_mask2_1_end_mask_0 = const()[name = string("pad_mask2_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<bool, [1, 4]> pad_mask2_1 = slice_by_index(begin = pad_mask2_1_begin_0, end = pad_mask2_1_end_0, end_mask = pad_mask2_1_end_mask_0, x = pad_mask1_1)[name = string("pad_mask2_1")];
            tensor<int32, [3]> mask_2_begin_0 = const()[name = string("mask_2_begin_0"), val = tensor<int32, [3]>([0, 56, 0])];
            tensor<int32, [3]> mask_2_end_0 = const()[name = string("mask_2_end_0"), val = tensor<int32, [3]>([1, 60, 60])];
            tensor<bool, [3]> mask_2_end_mask_0 = const()[name = string("mask_2_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<bool, [1, 4, 60]> mask_2 = slice_by_index(begin = mask_2_begin_0, end = mask_2_end_0, end_mask = mask_2_end_mask_0, x = att_mask4_1)[name = string("mask_2")];
            tensor<int32, [4]> cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 56, 1024])];
            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            string cache_last_channel_to_fp16_dtype_0 = const()[name = string("cache_last_channel_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [24, 1, 56, 1024]> cache_last_channel_to_fp16 = cast(dtype = cache_last_channel_to_fp16_dtype_0, x = cache_last_channel)[name = string("cast_7")];
            tensor<fp16, [1, 56, 1024]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache_1_cast_fp16")];
            tensor<int32, [4]> cache0_1_begin_0 = const()[name = string("cache0_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> cache0_1_end_0 = const()[name = string("cache0_1_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
            tensor<bool, [4]> cache0_1_end_mask_0 = const()[name = string("cache0_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache0_1_squeeze_mask_0 = const()[name = string("cache0_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            string cache_last_time_to_fp16_dtype_0 = const()[name = string("cache_last_time_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [24, 1, 1024, 8]> cache_last_time_to_fp16 = cast(dtype = cache_last_time_to_fp16_dtype_0, x = cache_last_time)[name = string("cast_6")];
            tensor<fp16, [1, 1024, 8]> cache0_1_cast_fp16 = slice_by_index(begin = cache0_1_begin_0, end = cache0_1_end_0, end_mask = cache0_1_end_mask_0, squeeze_mask = cache0_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache0_1_cast_fp16")];
            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604416)))];
            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606528)))];
            fp16 var_29_to_fp16 = const()[name = string("op_29_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 4, 1024]> input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_344_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608640))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8803008))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [4096]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8803584)))];
            tensor<fp16, [1, 4, 4096]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_450_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("op_450_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8811840))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13006208))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1024]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13006784)))];
            tensor<fp16, [1, 4, 1024]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_palettized, x = var_450_cast_fp16)[name = string("linear_2_cast_fp16")];
            fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_13_cast_fp16 = add(x = var_344_cast_fp16, y = var_456_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<int32, [1]> key_2_axes_0 = const()[name = string("key_2_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13008896)))];
            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13011008)))];
            tensor<fp16, [1, 4, 1024]> key_2_cast_fp16 = layer_norm(axes = key_2_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_13_cast_fp16)[name = string("key_2_cast_fp16")];
            bool input_15_interleave_0 = const()[name = string("input_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_15_cast_fp16 = concat(axis = var_64, interleave = input_15_interleave_0, values = (cache_1_cast_fp16, key_2_cast_fp16))[name = string("input_15_cast_fp16")];
            tensor<int32, [3]> var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")];
            bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_484_cast_fp16 = concat(axis = var_64, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_2_cast_fp16))[name = string("op_484_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13013120))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14061760))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_palettized, x = key_2_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_488 = const()[name = string("op_488"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_2_cast_fp16 = reshape(shape = var_488, x = linear_3_cast_fp16)[name = string("q_2_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14062336))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15110976))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [4]> var_492 = const()[name = string("op_492"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_2_cast_fp16 = reshape(shape = var_492, x = linear_4_cast_fp16)[name = string("k_2_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15111552))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16160192))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_5_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<int32, [4]> var_496 = const()[name = string("op_496"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_2_cast_fp16 = reshape(shape = var_496, x = linear_5_cast_fp16)[name = string("v_2_cast_fp16")];
            tensor<int32, [4]> value_4_perm_0 = const()[name = string("value_4_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16160768)))];
            tensor<fp16, [1, 4, 8, 128]> var_508_cast_fp16 = add(x = q_2_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_508_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16162880)))];
            tensor<fp16, [1, 4, 8, 128]> var_510_cast_fp16 = add(x = q_2_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_510_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_2_perm_0 = const()[name = string("q_with_bias_v_2_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_11_transpose_x_0 = const()[name = string("x_11_transpose_x_0"), val = bool(false)];
            bool x_11_transpose_y_0 = const()[name = string("x_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_512_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16164992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286912))))[name = string("op_512_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_2_cast_fp16 = transpose(perm = q_with_bias_v_2_perm_0, x = var_510_cast_fp16)[name = string("transpose_359")];
            tensor<fp16, [1, 8, 4, 119]> x_11_cast_fp16 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = q_with_bias_v_2_cast_fp16, y = op_512_to_fp16_palettized)[name = string("x_11_cast_fp16")];
            tensor<int32, [8]> x0_4_pad_0 = const()[name = string("x0_4_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_4_mode_0 = const()[name = string("x0_4_mode_0"), val = string("constant")];
            fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_4_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x0_4_mode_0, pad = x0_4_pad_0, x = x_11_cast_fp16)[name = string("x0_4_cast_fp16")];
            tensor<int32, [4]> var_520 = const()[name = string("op_520"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_2_cast_fp16 = reshape(shape = var_520, x = x0_4_cast_fp16)[name = string("x1_2_cast_fp16")];
            tensor<int32, [4]> var_524_begin_0 = const()[name = string("op_524_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_524_end_0 = const()[name = string("op_524_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_524_end_mask_0 = const()[name = string("op_524_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_524_cast_fp16 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = x1_2_cast_fp16)[name = string("op_524_cast_fp16")];
            tensor<int32, [4]> var_525 = const()[name = string("op_525"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_2_cast_fp16 = reshape(shape = var_525, x = var_524_cast_fp16)[name = string("matrix_bd_2_cast_fp16")];
            bool matrix_ac_2_transpose_x_0 = const()[name = string("matrix_ac_2_transpose_x_0"), val = bool(false)];
            bool matrix_ac_2_transpose_y_0 = const()[name = string("matrix_ac_2_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_2_cast_fp16)[name = string("transpose_357")];
            tensor<fp16, [1, 8, 4, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_508_cast_fp16)[name = string("transpose_358")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_2_cast_fp16 = matmul(transpose_x = matrix_ac_2_transpose_x_0, transpose_y = matrix_ac_2_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_2_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_2_begin_0 = const()[name = string("matrix_bd0_2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_2_end_0 = const()[name = string("matrix_bd0_2_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_2_end_mask_0 = const()[name = string("matrix_bd0_2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_2_cast_fp16 = slice_by_index(begin = matrix_bd0_2_begin_0, end = matrix_bd0_2_end_0, end_mask = matrix_bd0_2_end_mask_0, x = matrix_bd_2_cast_fp16)[name = string("matrix_bd0_2_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_534_cast_fp16 = add(x = matrix_ac_2_cast_fp16, y = matrix_bd0_2_cast_fp16)[name = string("op_534_cast_fp16")];
            fp16 _inversed_scores_2_y_0_to_fp16 = const()[name = string("_inversed_scores_2_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_2_cast_fp16 = mul(x = var_534_cast_fp16, y = _inversed_scores_2_y_0_to_fp16)[name = string("_inversed_scores_2_cast_fp16")];
            tensor<int32, [1]> mask0_4_axes_0 = const()[name = string("mask0_4_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 4, 60]> mask0_4 = expand_dims(axes = mask0_4_axes_0, x = mask_2)[name = string("mask0_4")];
            fp16 var_31_to_fp16 = const()[name = string("op_31_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 8, 4, 60]> scores0_2_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_2_cast_fp16, cond = mask0_4)[name = string("scores0_2_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_540_cast_fp16 = softmax(axis = var_56, x = scores0_2_cast_fp16)[name = string("op_540_cast_fp16")];
            fp16 var_30_to_fp16 = const()[name = string("op_30_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 60]> input0_13_cast_fp16 = select(a = var_30_to_fp16, b = var_540_cast_fp16, cond = mask0_4)[name = string("input0_13_cast_fp16")];
            bool x2_2_transpose_x_0 = const()[name = string("x2_2_transpose_x_0"), val = bool(false)];
            bool x2_2_transpose_y_0 = const()[name = string("x2_2_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_4_cast_fp16 = transpose(perm = value_4_perm_0, x = v_2_cast_fp16)[name = string("transpose_356")];
            tensor<fp16, [1, 8, 4, 128]> x2_2_cast_fp16 = matmul(transpose_x = x2_2_transpose_x_0, transpose_y = x2_2_transpose_y_0, x = input0_13_cast_fp16, y = value_4_cast_fp16)[name = string("x2_2_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_545 = const()[name = string("op_545"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = x2_2_cast_fp16)[name = string("transpose_355")];
            tensor<fp16, [1, 4, 1024]> input1_8_cast_fp16 = reshape(shape = var_545, x = var_544_cast_fp16)[name = string("input1_8_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16287488))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17336128))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_palettized, x = input1_8_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_7_cast_fp16)[name = string("input0_19_cast_fp16")];
            tensor<int32, [1]> x_15_axes_0 = const()[name = string("x_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17336704)))];
            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17338816)))];
            tensor<fp16, [1, 4, 1024]> x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input0_19_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [3]> input_17_perm_0 = const()[name = string("input_17_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_15_pad_type_0 = const()[name = string("input0_15_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_15_strides_0 = const()[name = string("input0_15_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_15_pad_0 = const()[name = string("input0_15_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_15_dilations_0 = const()[name = string("input0_15_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_15_groups_0 = const()[name = string("input0_15_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17340928))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19438144))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_17_cast_fp16 = transpose(perm = input_17_perm_0, x = x_15_cast_fp16)[name = string("transpose_354")];
            tensor<fp16, [1, 2048, 4]> input0_15_cast_fp16 = conv(dilations = input0_15_dilations_0, groups = input0_15_groups_0, pad = input0_15_pad_0, pad_type = input0_15_pad_type_0, strides = input0_15_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = string("input0_15_cast_fp16")];
            int32 x_17_split_num_splits_0 = const()[name = string("x_17_split_num_splits_0"), val = int32(2)];
            int32 x_17_split_axis_0 = const()[name = string("x_17_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_17_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_17_split_cast_fp16_1 = split(axis = x_17_split_axis_0, num_splits = x_17_split_num_splits_0, x = input0_15_cast_fp16)[name = string("x_17_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_17_split_1_sigmoid_cast_fp16 = sigmoid(x = x_17_split_cast_fp16_1)[name = string("x_17_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_17_cast_fp16 = mul(x = x_17_split_cast_fp16_0, y = x_17_split_1_sigmoid_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<int32, [1]> var_570_axes_0 = const()[name = string("op_570_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 4]> var_570 = expand_dims(axes = var_570_axes_0, x = pad_mask2_1)[name = string("op_570")];
            tensor<fp16, [1, 1024, 4]> input3_4_cast_fp16 = select(a = var_30_to_fp16, b = x_17_cast_fp16, cond = var_570)[name = string("input3_4_cast_fp16")];
            bool new_x0_2_interleave_0 = const()[name = string("new_x0_2_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_2_cast_fp16 = concat(axis = var_56, interleave = new_x0_2_interleave_0, values = (cache0_1_cast_fp16, input3_4_cast_fp16))[name = string("new_x0_2_cast_fp16")];
            tensor<int32, [3]> var_583_begin_0 = const()[name = string("op_583_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_583_end_0 = const()[name = string("op_583_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_583_end_mask_0 = const()[name = string("op_583_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_583_cast_fp16 = slice_by_index(begin = var_583_begin_0, end = var_583_end_0, end_mask = var_583_end_mask_0, x = new_x0_2_cast_fp16)[name = string("op_583_cast_fp16")];
            string x_19_pad_type_0 = const()[name = string("x_19_pad_type_0"), val = string("valid")];
            int32 x_19_groups_0 = const()[name = string("x_19_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_19_strides_0 = const()[name = string("x_19_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_19_pad_0 = const()[name = string("x_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_19_dilations_0 = const()[name = string("x_19_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19438720))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19448000))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_19_cast_fp16 = conv(dilations = x_19_dilations_0, groups = x_19_groups_0, pad = x_19_pad_0, pad_type = x_19_pad_type_0, strides = x_19_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_2_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<int32, [3]> input4_1_perm_0 = const()[name = string("input4_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_21_axes_0 = const()[name = string("x_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19448576)))];
            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19450688)))];
            tensor<fp16, [1, 4, 1024]> input4_1_cast_fp16 = transpose(perm = input4_1_perm_0, x = x_19_cast_fp16)[name = string("transpose_353")];
            tensor<fp16, [1, 4, 1024]> x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input4_1_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [3]> input5_1_perm_0 = const()[name = string("input5_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input5_1_cast_fp16 = transpose(perm = input5_1_perm_0, x = x_21_cast_fp16)[name = string("transpose_352")];
            tensor<fp16, [1, 1024, 4]> var_598_cast_fp16 = silu(x = input5_1_cast_fp16)[name = string("op_598_cast_fp16")];
            string x_23_pad_type_0 = const()[name = string("x_23_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_23_strides_0 = const()[name = string("x_23_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_23_pad_0 = const()[name = string("x_23_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_23_dilations_0 = const()[name = string("x_23_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_23_groups_0 = const()[name = string("x_23_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19452800))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20501440))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_23_cast_fp16 = conv(dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_598_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<int32, [3]> input6_1_perm_0 = const()[name = string("input6_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input6_1_cast_fp16 = transpose(perm = input6_1_perm_0, x = x_23_cast_fp16)[name = string("transpose_351")];
            tensor<fp16, [1, 4, 1024]> input1_10_cast_fp16 = add(x = input0_19_cast_fp16, y = input6_1_cast_fp16)[name = string("input1_10_cast_fp16")];
            tensor<int32, [1]> input0_17_axes_0 = const()[name = string("input0_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20502016)))];
            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20504128)))];
            tensor<fp16, [1, 4, 1024]> input0_17_cast_fp16 = layer_norm(axes = input0_17_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input1_10_cast_fp16)[name = string("input0_17_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20506240))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24700608))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_17_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_619_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("op_619_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24701184))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28895552))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_palettized, x = var_619_cast_fp16)[name = string("linear_9_cast_fp16")];
            fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_625_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_624_to_fp16)[name = string("op_625_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_4_cast_fp16 = add(x = input1_10_cast_fp16, y = var_625_cast_fp16)[name = string("input2_4_cast_fp16")];
            tensor<int32, [1]> input0_21_axes_0 = const()[name = string("input0_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28896128)))];
            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28898240)))];
            tensor<fp16, [1, 4, 1024]> input0_21_cast_fp16 = layer_norm(axes = input0_21_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input2_4_cast_fp16)[name = string("input0_21_cast_fp16")];
            tensor<int32, [4]> cache1_1_begin_0 = const()[name = string("cache1_1_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> cache1_1_end_0 = const()[name = string("cache1_1_end_0"), val = tensor<int32, [4]>([2, 1, 56, 1024])];
            tensor<bool, [4]> cache1_1_end_mask_0 = const()[name = string("cache1_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache1_1_squeeze_mask_0 = const()[name = string("cache1_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache1_1_cast_fp16 = slice_by_index(begin = cache1_1_begin_0, end = cache1_1_end_0, end_mask = cache1_1_end_mask_0, squeeze_mask = cache1_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache1_1_cast_fp16")];
            tensor<int32, [4]> cache2_1_begin_0 = const()[name = string("cache2_1_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> cache2_1_end_0 = const()[name = string("cache2_1_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
            tensor<bool, [4]> cache2_1_end_mask_0 = const()[name = string("cache2_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache2_1_squeeze_mask_0 = const()[name = string("cache2_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache2_1_cast_fp16 = slice_by_index(begin = cache2_1_begin_0, end = cache2_1_end_0, end_mask = cache2_1_end_mask_0, squeeze_mask = cache2_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache2_1_cast_fp16")];
            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28900352)))];
            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28902464)))];
            tensor<fp16, [1, 4, 1024]> input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input0_21_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28904576))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33098944))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_654_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("op_654_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33099520))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37293888))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_palettized, x = var_654_cast_fp16)[name = string("linear_11_cast_fp16")];
            fp16 var_659_to_fp16 = const()[name = string("op_659_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_660_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_659_to_fp16)[name = string("op_660_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_25_cast_fp16 = add(x = input0_21_cast_fp16, y = var_660_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<int32, [1]> key_4_axes_0 = const()[name = string("key_4_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37294464)))];
            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37296576)))];
            tensor<fp16, [1, 4, 1024]> key_4_cast_fp16 = layer_norm(axes = key_4_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_25_cast_fp16)[name = string("key_4_cast_fp16")];
            bool input_27_interleave_0 = const()[name = string("input_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_27_cast_fp16 = concat(axis = var_64, interleave = input_27_interleave_0, values = (cache1_1_cast_fp16, key_4_cast_fp16))[name = string("input_27_cast_fp16")];
            tensor<int32, [3]> var_682_begin_0 = const()[name = string("op_682_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_682_end_0 = const()[name = string("op_682_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_682_end_mask_0 = const()[name = string("op_682_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_682_cast_fp16 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = cache1_1_cast_fp16)[name = string("op_682_cast_fp16")];
            bool var_688_interleave_0 = const()[name = string("op_688_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_688_cast_fp16 = concat(axis = var_64, interleave = var_688_interleave_0, values = (var_682_cast_fp16, key_4_cast_fp16))[name = string("op_688_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37298688))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38347328))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_palettized, x = key_4_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_4_cast_fp16 = reshape(shape = var_692, x = linear_12_cast_fp16)[name = string("q_4_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38347904))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39396544))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<int32, [4]> var_696 = const()[name = string("op_696"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_4_cast_fp16 = reshape(shape = var_696, x = linear_13_cast_fp16)[name = string("k_4_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39397120))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40445760))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_700 = const()[name = string("op_700"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_4_cast_fp16 = reshape(shape = var_700, x = linear_14_cast_fp16)[name = string("v_4_cast_fp16")];
            tensor<int32, [4]> value_6_perm_0 = const()[name = string("value_6_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40446336)))];
            tensor<fp16, [1, 4, 8, 128]> var_712_cast_fp16 = add(x = q_4_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_712_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40448448)))];
            tensor<fp16, [1, 4, 8, 128]> var_714_cast_fp16 = add(x = q_4_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_714_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_4_perm_0 = const()[name = string("q_with_bias_v_4_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_31_transpose_x_0 = const()[name = string("x_31_transpose_x_0"), val = bool(false)];
            bool x_31_transpose_y_0 = const()[name = string("x_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_716_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40450560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40572480))))[name = string("op_716_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_4_cast_fp16 = transpose(perm = q_with_bias_v_4_perm_0, x = var_714_cast_fp16)[name = string("transpose_350")];
            tensor<fp16, [1, 8, 4, 119]> x_31_cast_fp16 = matmul(transpose_x = x_31_transpose_x_0, transpose_y = x_31_transpose_y_0, x = q_with_bias_v_4_cast_fp16, y = op_716_to_fp16_palettized)[name = string("x_31_cast_fp16")];
            tensor<int32, [8]> x0_6_pad_0 = const()[name = string("x0_6_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_6_mode_0 = const()[name = string("x0_6_mode_0"), val = string("constant")];
            fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_6_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x0_6_mode_0, pad = x0_6_pad_0, x = x_31_cast_fp16)[name = string("x0_6_cast_fp16")];
            tensor<int32, [4]> var_724 = const()[name = string("op_724"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_4_cast_fp16 = reshape(shape = var_724, x = x0_6_cast_fp16)[name = string("x1_4_cast_fp16")];
            tensor<int32, [4]> var_728_begin_0 = const()[name = string("op_728_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_728_end_0 = const()[name = string("op_728_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_728_end_mask_0 = const()[name = string("op_728_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_728_cast_fp16 = slice_by_index(begin = var_728_begin_0, end = var_728_end_0, end_mask = var_728_end_mask_0, x = x1_4_cast_fp16)[name = string("op_728_cast_fp16")];
            tensor<int32, [4]> var_729 = const()[name = string("op_729"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_4_cast_fp16 = reshape(shape = var_729, x = var_728_cast_fp16)[name = string("matrix_bd_4_cast_fp16")];
            bool matrix_ac_4_transpose_x_0 = const()[name = string("matrix_ac_4_transpose_x_0"), val = bool(false)];
            bool matrix_ac_4_transpose_y_0 = const()[name = string("matrix_ac_4_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_4_cast_fp16)[name = string("transpose_348")];
            tensor<fp16, [1, 8, 4, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_712_cast_fp16)[name = string("transpose_349")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_4_cast_fp16 = matmul(transpose_x = matrix_ac_4_transpose_x_0, transpose_y = matrix_ac_4_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_4_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_4_begin_0 = const()[name = string("matrix_bd0_4_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_4_end_0 = const()[name = string("matrix_bd0_4_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_4_end_mask_0 = const()[name = string("matrix_bd0_4_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_4_cast_fp16 = slice_by_index(begin = matrix_bd0_4_begin_0, end = matrix_bd0_4_end_0, end_mask = matrix_bd0_4_end_mask_0, x = matrix_bd_4_cast_fp16)[name = string("matrix_bd0_4_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_738_cast_fp16 = add(x = matrix_ac_4_cast_fp16, y = matrix_bd0_4_cast_fp16)[name = string("op_738_cast_fp16")];
            fp16 _inversed_scores_4_y_0_to_fp16 = const()[name = string("_inversed_scores_4_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_4_cast_fp16 = mul(x = var_738_cast_fp16, y = _inversed_scores_4_y_0_to_fp16)[name = string("_inversed_scores_4_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_4_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_4_cast_fp16, cond = mask0_4)[name = string("scores0_4_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_744_cast_fp16 = softmax(axis = var_56, x = scores0_4_cast_fp16)[name = string("op_744_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_23_cast_fp16 = select(a = var_30_to_fp16, b = var_744_cast_fp16, cond = mask0_4)[name = string("input0_23_cast_fp16")];
            bool x2_4_transpose_x_0 = const()[name = string("x2_4_transpose_x_0"), val = bool(false)];
            bool x2_4_transpose_y_0 = const()[name = string("x2_4_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_6_cast_fp16 = transpose(perm = value_6_perm_0, x = v_4_cast_fp16)[name = string("transpose_347")];
            tensor<fp16, [1, 8, 4, 128]> x2_4_cast_fp16 = matmul(transpose_x = x2_4_transpose_x_0, transpose_y = x2_4_transpose_y_0, x = input0_23_cast_fp16, y = value_6_cast_fp16)[name = string("x2_4_cast_fp16")];
            tensor<int32, [4]> var_748_perm_0 = const()[name = string("op_748_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_749 = const()[name = string("op_749"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_748_cast_fp16 = transpose(perm = var_748_perm_0, x = x2_4_cast_fp16)[name = string("transpose_346")];
            tensor<fp16, [1, 4, 1024]> input1_12_cast_fp16 = reshape(shape = var_749, x = var_748_cast_fp16)[name = string("input1_12_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40573056))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41621696))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_palettized, x = input1_12_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_25_cast_fp16 = add(x = input_25_cast_fp16, y = linear_16_cast_fp16)[name = string("input0_25_cast_fp16")];
            tensor<int32, [1]> x_35_axes_0 = const()[name = string("x_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41622272)))];
            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41624384)))];
            tensor<fp16, [1, 4, 1024]> x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input0_25_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> input_29_perm_0 = const()[name = string("input_29_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_27_pad_type_0 = const()[name = string("input0_27_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_27_strides_0 = const()[name = string("input0_27_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_27_pad_0 = const()[name = string("input0_27_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_27_dilations_0 = const()[name = string("input0_27_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_27_groups_0 = const()[name = string("input0_27_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41626496))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43723712))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_29_cast_fp16 = transpose(perm = input_29_perm_0, x = x_35_cast_fp16)[name = string("transpose_345")];
            tensor<fp16, [1, 2048, 4]> input0_27_cast_fp16 = conv(dilations = input0_27_dilations_0, groups = input0_27_groups_0, pad = input0_27_pad_0, pad_type = input0_27_pad_type_0, strides = input0_27_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("input0_27_cast_fp16")];
            int32 x_37_split_num_splits_0 = const()[name = string("x_37_split_num_splits_0"), val = int32(2)];
            int32 x_37_split_axis_0 = const()[name = string("x_37_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_37_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_37_split_cast_fp16_1 = split(axis = x_37_split_axis_0, num_splits = x_37_split_num_splits_0, x = input0_27_cast_fp16)[name = string("x_37_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_37_split_1_sigmoid_cast_fp16 = sigmoid(x = x_37_split_cast_fp16_1)[name = string("x_37_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_37_cast_fp16 = mul(x = x_37_split_cast_fp16_0, y = x_37_split_1_sigmoid_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_29_cast_fp16 = select(a = var_30_to_fp16, b = x_37_cast_fp16, cond = var_570)[name = string("input0_29_cast_fp16")];
            bool new_x0_4_interleave_0 = const()[name = string("new_x0_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_4_cast_fp16 = concat(axis = var_56, interleave = new_x0_4_interleave_0, values = (cache2_1_cast_fp16, input0_29_cast_fp16))[name = string("new_x0_4_cast_fp16")];
            tensor<int32, [3]> var_787_begin_0 = const()[name = string("op_787_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_787_end_0 = const()[name = string("op_787_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_787_end_mask_0 = const()[name = string("op_787_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_787_cast_fp16 = slice_by_index(begin = var_787_begin_0, end = var_787_end_0, end_mask = var_787_end_mask_0, x = new_x0_4_cast_fp16)[name = string("op_787_cast_fp16")];
            string x_39_pad_type_0 = const()[name = string("x_39_pad_type_0"), val = string("valid")];
            int32 x_39_groups_0 = const()[name = string("x_39_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_39_strides_0 = const()[name = string("x_39_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_39_pad_0 = const()[name = string("x_39_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_39_dilations_0 = const()[name = string("x_39_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43724288))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43733568))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_39_cast_fp16 = conv(dilations = x_39_dilations_0, groups = x_39_groups_0, pad = x_39_pad_0, pad_type = x_39_pad_type_0, strides = x_39_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_4_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [3]> input1_14_perm_0 = const()[name = string("input1_14_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_41_axes_0 = const()[name = string("x_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43734144)))];
            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43736256)))];
            tensor<fp16, [1, 4, 1024]> input1_14_cast_fp16 = transpose(perm = input1_14_perm_0, x = x_39_cast_fp16)[name = string("transpose_344")];
            tensor<fp16, [1, 4, 1024]> x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input1_14_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [3]> input2_8_perm_0 = const()[name = string("input2_8_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_8_cast_fp16 = transpose(perm = input2_8_perm_0, x = x_41_cast_fp16)[name = string("transpose_343")];
            tensor<fp16, [1, 1024, 4]> var_802_cast_fp16 = silu(x = input2_8_cast_fp16)[name = string("op_802_cast_fp16")];
            string x_43_pad_type_0 = const()[name = string("x_43_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_43_strides_0 = const()[name = string("x_43_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_43_pad_0 = const()[name = string("x_43_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_43_dilations_0 = const()[name = string("x_43_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_43_groups_0 = const()[name = string("x_43_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43738368))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44787008))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_43_cast_fp16 = conv(dilations = x_43_dilations_0, groups = x_43_groups_0, pad = x_43_pad_0, pad_type = x_43_pad_type_0, strides = x_43_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_802_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<int32, [3]> input3_6_perm_0 = const()[name = string("input3_6_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_6_cast_fp16 = transpose(perm = input3_6_perm_0, x = x_43_cast_fp16)[name = string("transpose_342")];
            tensor<fp16, [1, 4, 1024]> input1_16_cast_fp16 = add(x = input0_25_cast_fp16, y = input3_6_cast_fp16)[name = string("input1_16_cast_fp16")];
            tensor<int32, [1]> input0_31_axes_0 = const()[name = string("input0_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44787584)))];
            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44789696)))];
            tensor<fp16, [1, 4, 1024]> input0_31_cast_fp16 = layer_norm(axes = input0_31_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input1_16_cast_fp16)[name = string("input0_31_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44791808))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48986176))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_31_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_823_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("op_823_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48986752))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53181120))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_palettized, x = var_823_cast_fp16)[name = string("linear_18_cast_fp16")];
            fp16 var_828_to_fp16 = const()[name = string("op_828_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_829_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_828_to_fp16)[name = string("op_829_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_10_cast_fp16 = add(x = input1_16_cast_fp16, y = var_829_cast_fp16)[name = string("input2_10_cast_fp16")];
            tensor<int32, [1]> input0_33_axes_0 = const()[name = string("input0_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53181696)))];
            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53183808)))];
            tensor<fp16, [1, 4, 1024]> input0_33_cast_fp16 = layer_norm(axes = input0_33_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input2_10_cast_fp16)[name = string("input0_33_cast_fp16")];
            tensor<int32, [4]> cache3_1_begin_0 = const()[name = string("cache3_1_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> cache3_1_end_0 = const()[name = string("cache3_1_end_0"), val = tensor<int32, [4]>([3, 1, 56, 1024])];
            tensor<bool, [4]> cache3_1_end_mask_0 = const()[name = string("cache3_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache3_1_squeeze_mask_0 = const()[name = string("cache3_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache3_1_cast_fp16 = slice_by_index(begin = cache3_1_begin_0, end = cache3_1_end_0, end_mask = cache3_1_end_mask_0, squeeze_mask = cache3_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache3_1_cast_fp16")];
            tensor<int32, [4]> cache4_1_begin_0 = const()[name = string("cache4_1_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> cache4_1_end_0 = const()[name = string("cache4_1_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
            tensor<bool, [4]> cache4_1_end_mask_0 = const()[name = string("cache4_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache4_1_squeeze_mask_0 = const()[name = string("cache4_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache4_1_cast_fp16 = slice_by_index(begin = cache4_1_begin_0, end = cache4_1_end_0, end_mask = cache4_1_end_mask_0, squeeze_mask = cache4_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache4_1_cast_fp16")];
            tensor<int32, [1]> input_33_axes_0 = const()[name = string("input_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53185920)))];
            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53188032)))];
            tensor<fp16, [1, 4, 1024]> input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input0_33_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53190144))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57384512))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_858_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("op_858_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57385088))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61579456))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_palettized, x = var_858_cast_fp16)[name = string("linear_20_cast_fp16")];
            fp16 var_863_to_fp16 = const()[name = string("op_863_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_864_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_863_to_fp16)[name = string("op_864_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_37_cast_fp16 = add(x = input0_33_cast_fp16, y = var_864_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<int32, [1]> key_6_axes_0 = const()[name = string("key_6_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61580032)))];
            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61582144)))];
            tensor<fp16, [1, 4, 1024]> key_6_cast_fp16 = layer_norm(axes = key_6_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_6_cast_fp16")];
            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_39_cast_fp16 = concat(axis = var_64, interleave = input_39_interleave_0, values = (cache3_1_cast_fp16, key_6_cast_fp16))[name = string("input_39_cast_fp16")];
            tensor<int32, [3]> var_886_begin_0 = const()[name = string("op_886_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_886_end_0 = const()[name = string("op_886_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_886_end_mask_0 = const()[name = string("op_886_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_886_cast_fp16 = slice_by_index(begin = var_886_begin_0, end = var_886_end_0, end_mask = var_886_end_mask_0, x = cache3_1_cast_fp16)[name = string("op_886_cast_fp16")];
            bool var_892_interleave_0 = const()[name = string("op_892_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_892_cast_fp16 = concat(axis = var_64, interleave = var_892_interleave_0, values = (var_886_cast_fp16, key_6_cast_fp16))[name = string("op_892_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61584256))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62632896))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_palettized, x = key_6_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_896 = const()[name = string("op_896"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_6_cast_fp16 = reshape(shape = var_896, x = linear_21_cast_fp16)[name = string("q_6_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62633472))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63682112))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_900 = const()[name = string("op_900"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_6_cast_fp16 = reshape(shape = var_900, x = linear_22_cast_fp16)[name = string("k_6_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63682688))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64731328))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_904 = const()[name = string("op_904"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_6_cast_fp16 = reshape(shape = var_904, x = linear_23_cast_fp16)[name = string("v_6_cast_fp16")];
            tensor<int32, [4]> value_8_perm_0 = const()[name = string("value_8_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64731904)))];
            tensor<fp16, [1, 4, 8, 128]> var_916_cast_fp16 = add(x = q_6_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_916_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64734016)))];
            tensor<fp16, [1, 4, 8, 128]> var_918_cast_fp16 = add(x = q_6_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_918_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_6_perm_0 = const()[name = string("q_with_bias_v_6_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_51_transpose_x_0 = const()[name = string("x_51_transpose_x_0"), val = bool(false)];
            bool x_51_transpose_y_0 = const()[name = string("x_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_920_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64736128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64858048))))[name = string("op_920_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_6_cast_fp16 = transpose(perm = q_with_bias_v_6_perm_0, x = var_918_cast_fp16)[name = string("transpose_341")];
            tensor<fp16, [1, 8, 4, 119]> x_51_cast_fp16 = matmul(transpose_x = x_51_transpose_x_0, transpose_y = x_51_transpose_y_0, x = q_with_bias_v_6_cast_fp16, y = op_920_to_fp16_palettized)[name = string("x_51_cast_fp16")];
            tensor<int32, [8]> x0_8_pad_0 = const()[name = string("x0_8_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_8_mode_0 = const()[name = string("x0_8_mode_0"), val = string("constant")];
            fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_8_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x0_8_mode_0, pad = x0_8_pad_0, x = x_51_cast_fp16)[name = string("x0_8_cast_fp16")];
            tensor<int32, [4]> var_928 = const()[name = string("op_928"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_6_cast_fp16 = reshape(shape = var_928, x = x0_8_cast_fp16)[name = string("x1_6_cast_fp16")];
            tensor<int32, [4]> var_932_begin_0 = const()[name = string("op_932_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_932_end_0 = const()[name = string("op_932_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_932_end_mask_0 = const()[name = string("op_932_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_932_cast_fp16 = slice_by_index(begin = var_932_begin_0, end = var_932_end_0, end_mask = var_932_end_mask_0, x = x1_6_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<int32, [4]> var_933 = const()[name = string("op_933"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_6_cast_fp16 = reshape(shape = var_933, x = var_932_cast_fp16)[name = string("matrix_bd_6_cast_fp16")];
            bool matrix_ac_6_transpose_x_0 = const()[name = string("matrix_ac_6_transpose_x_0"), val = bool(false)];
            bool matrix_ac_6_transpose_y_0 = const()[name = string("matrix_ac_6_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_6_cast_fp16)[name = string("transpose_339")];
            tensor<fp16, [1, 8, 4, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_916_cast_fp16)[name = string("transpose_340")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_6_cast_fp16 = matmul(transpose_x = matrix_ac_6_transpose_x_0, transpose_y = matrix_ac_6_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_6_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_6_begin_0 = const()[name = string("matrix_bd0_6_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_6_end_0 = const()[name = string("matrix_bd0_6_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_6_end_mask_0 = const()[name = string("matrix_bd0_6_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_6_cast_fp16 = slice_by_index(begin = matrix_bd0_6_begin_0, end = matrix_bd0_6_end_0, end_mask = matrix_bd0_6_end_mask_0, x = matrix_bd_6_cast_fp16)[name = string("matrix_bd0_6_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_942_cast_fp16 = add(x = matrix_ac_6_cast_fp16, y = matrix_bd0_6_cast_fp16)[name = string("op_942_cast_fp16")];
            fp16 _inversed_scores_6_y_0_to_fp16 = const()[name = string("_inversed_scores_6_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_6_cast_fp16 = mul(x = var_942_cast_fp16, y = _inversed_scores_6_y_0_to_fp16)[name = string("_inversed_scores_6_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_6_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_6_cast_fp16, cond = mask0_4)[name = string("scores0_6_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_948_cast_fp16 = softmax(axis = var_56, x = scores0_6_cast_fp16)[name = string("op_948_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_35_cast_fp16 = select(a = var_30_to_fp16, b = var_948_cast_fp16, cond = mask0_4)[name = string("input0_35_cast_fp16")];
            bool x2_6_transpose_x_0 = const()[name = string("x2_6_transpose_x_0"), val = bool(false)];
            bool x2_6_transpose_y_0 = const()[name = string("x2_6_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_8_cast_fp16 = transpose(perm = value_8_perm_0, x = v_6_cast_fp16)[name = string("transpose_338")];
            tensor<fp16, [1, 8, 4, 128]> x2_6_cast_fp16 = matmul(transpose_x = x2_6_transpose_x_0, transpose_y = x2_6_transpose_y_0, x = input0_35_cast_fp16, y = value_8_cast_fp16)[name = string("x2_6_cast_fp16")];
            tensor<int32, [4]> var_952_perm_0 = const()[name = string("op_952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_953 = const()[name = string("op_953"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_952_cast_fp16 = transpose(perm = var_952_perm_0, x = x2_6_cast_fp16)[name = string("transpose_337")];
            tensor<fp16, [1, 4, 1024]> input1_18_cast_fp16 = reshape(shape = var_953, x = var_952_cast_fp16)[name = string("input1_18_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64858624))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65907264))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input1_18_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_37_cast_fp16 = add(x = input_37_cast_fp16, y = linear_25_cast_fp16)[name = string("input0_37_cast_fp16")];
            tensor<int32, [1]> x_55_axes_0 = const()[name = string("x_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65907840)))];
            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65909952)))];
            tensor<fp16, [1, 4, 1024]> x_55_cast_fp16 = layer_norm(axes = x_55_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input0_37_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<int32, [3]> input_41_perm_0 = const()[name = string("input_41_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_39_pad_type_0 = const()[name = string("input0_39_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_39_strides_0 = const()[name = string("input0_39_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_39_pad_0 = const()[name = string("input0_39_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_39_dilations_0 = const()[name = string("input0_39_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_39_groups_0 = const()[name = string("input0_39_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65912064))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68009280))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_41_cast_fp16 = transpose(perm = input_41_perm_0, x = x_55_cast_fp16)[name = string("transpose_336")];
            tensor<fp16, [1, 2048, 4]> input0_39_cast_fp16 = conv(dilations = input0_39_dilations_0, groups = input0_39_groups_0, pad = input0_39_pad_0, pad_type = input0_39_pad_type_0, strides = input0_39_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = string("input0_39_cast_fp16")];
            int32 x_57_split_num_splits_0 = const()[name = string("x_57_split_num_splits_0"), val = int32(2)];
            int32 x_57_split_axis_0 = const()[name = string("x_57_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_57_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_57_split_cast_fp16_1 = split(axis = x_57_split_axis_0, num_splits = x_57_split_num_splits_0, x = input0_39_cast_fp16)[name = string("x_57_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_57_split_1_sigmoid_cast_fp16 = sigmoid(x = x_57_split_cast_fp16_1)[name = string("x_57_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_57_cast_fp16 = mul(x = x_57_split_cast_fp16_0, y = x_57_split_1_sigmoid_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_41_cast_fp16 = select(a = var_30_to_fp16, b = x_57_cast_fp16, cond = var_570)[name = string("input0_41_cast_fp16")];
            bool new_x0_6_interleave_0 = const()[name = string("new_x0_6_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_6_cast_fp16 = concat(axis = var_56, interleave = new_x0_6_interleave_0, values = (cache4_1_cast_fp16, input0_41_cast_fp16))[name = string("new_x0_6_cast_fp16")];
            tensor<int32, [3]> var_991_begin_0 = const()[name = string("op_991_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_991_end_0 = const()[name = string("op_991_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_991_end_mask_0 = const()[name = string("op_991_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_991_cast_fp16 = slice_by_index(begin = var_991_begin_0, end = var_991_end_0, end_mask = var_991_end_mask_0, x = new_x0_6_cast_fp16)[name = string("op_991_cast_fp16")];
            string x_59_pad_type_0 = const()[name = string("x_59_pad_type_0"), val = string("valid")];
            int32 x_59_groups_0 = const()[name = string("x_59_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_59_strides_0 = const()[name = string("x_59_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_59_pad_0 = const()[name = string("x_59_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_59_dilations_0 = const()[name = string("x_59_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68009856))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68019136))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_59_cast_fp16 = conv(dilations = x_59_dilations_0, groups = x_59_groups_0, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = x_59_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_6_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [3]> input1_20_perm_0 = const()[name = string("input1_20_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_61_axes_0 = const()[name = string("x_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68019712)))];
            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68021824)))];
            tensor<fp16, [1, 4, 1024]> input1_20_cast_fp16 = transpose(perm = input1_20_perm_0, x = x_59_cast_fp16)[name = string("transpose_335")];
            tensor<fp16, [1, 4, 1024]> x_61_cast_fp16 = layer_norm(axes = x_61_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input1_20_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [3]> input2_12_perm_0 = const()[name = string("input2_12_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_12_cast_fp16 = transpose(perm = input2_12_perm_0, x = x_61_cast_fp16)[name = string("transpose_334")];
            tensor<fp16, [1, 1024, 4]> var_1006_cast_fp16 = silu(x = input2_12_cast_fp16)[name = string("op_1006_cast_fp16")];
            string x_63_pad_type_0 = const()[name = string("x_63_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_63_strides_0 = const()[name = string("x_63_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_63_pad_0 = const()[name = string("x_63_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_63_dilations_0 = const()[name = string("x_63_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_63_groups_0 = const()[name = string("x_63_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68023936))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69072576))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_63_cast_fp16 = conv(dilations = x_63_dilations_0, groups = x_63_groups_0, pad = x_63_pad_0, pad_type = x_63_pad_type_0, strides = x_63_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_1006_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<int32, [3]> input3_8_perm_0 = const()[name = string("input3_8_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_8_cast_fp16 = transpose(perm = input3_8_perm_0, x = x_63_cast_fp16)[name = string("transpose_333")];
            tensor<fp16, [1, 4, 1024]> input1_22_cast_fp16 = add(x = input0_37_cast_fp16, y = input3_8_cast_fp16)[name = string("input1_22_cast_fp16")];
            tensor<int32, [1]> input0_43_axes_0 = const()[name = string("input0_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69073152)))];
            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69075264)))];
            tensor<fp16, [1, 4, 1024]> input0_43_cast_fp16 = layer_norm(axes = input0_43_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input1_22_cast_fp16)[name = string("input0_43_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69077376))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73271744))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_43_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1027_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73272320))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77466688))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = var_1027_cast_fp16)[name = string("linear_27_cast_fp16")];
            fp16 var_1032_to_fp16 = const()[name = string("op_1032_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1033_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1032_to_fp16)[name = string("op_1033_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_14_cast_fp16 = add(x = input1_22_cast_fp16, y = var_1033_cast_fp16)[name = string("input2_14_cast_fp16")];
            tensor<int32, [1]> input0_45_axes_0 = const()[name = string("input0_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77467264)))];
            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77469376)))];
            tensor<fp16, [1, 4, 1024]> input0_45_cast_fp16 = layer_norm(axes = input0_45_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input2_14_cast_fp16)[name = string("input0_45_cast_fp16")];
            tensor<int32, [4]> cache5_1_begin_0 = const()[name = string("cache5_1_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> cache5_1_end_0 = const()[name = string("cache5_1_end_0"), val = tensor<int32, [4]>([4, 1, 56, 1024])];
            tensor<bool, [4]> cache5_1_end_mask_0 = const()[name = string("cache5_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache5_1_squeeze_mask_0 = const()[name = string("cache5_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache5_1_cast_fp16 = slice_by_index(begin = cache5_1_begin_0, end = cache5_1_end_0, end_mask = cache5_1_end_mask_0, squeeze_mask = cache5_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache5_1_cast_fp16")];
            tensor<int32, [4]> cache6_1_begin_0 = const()[name = string("cache6_1_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> cache6_1_end_0 = const()[name = string("cache6_1_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
            tensor<bool, [4]> cache6_1_end_mask_0 = const()[name = string("cache6_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache6_1_squeeze_mask_0 = const()[name = string("cache6_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache6_1_cast_fp16 = slice_by_index(begin = cache6_1_begin_0, end = cache6_1_end_0, end_mask = cache6_1_end_mask_0, squeeze_mask = cache6_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache6_1_cast_fp16")];
            tensor<int32, [1]> input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77471488)))];
            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77473600)))];
            tensor<fp16, [1, 4, 1024]> input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input0_45_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77475712))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81670080))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1062_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("op_1062_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81670656))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85865024))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = var_1062_cast_fp16)[name = string("linear_29_cast_fp16")];
            fp16 var_1067_to_fp16 = const()[name = string("op_1067_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1068_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1067_to_fp16)[name = string("op_1068_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_49_cast_fp16 = add(x = input0_45_cast_fp16, y = var_1068_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<int32, [1]> key_8_axes_0 = const()[name = string("key_8_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85865600)))];
            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85867712)))];
            tensor<fp16, [1, 4, 1024]> key_8_cast_fp16 = layer_norm(axes = key_8_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_49_cast_fp16)[name = string("key_8_cast_fp16")];
            bool input_51_interleave_0 = const()[name = string("input_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_51_cast_fp16 = concat(axis = var_64, interleave = input_51_interleave_0, values = (cache5_1_cast_fp16, key_8_cast_fp16))[name = string("input_51_cast_fp16")];
            tensor<int32, [3]> var_1090_begin_0 = const()[name = string("op_1090_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1090_end_0 = const()[name = string("op_1090_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_1090_end_mask_0 = const()[name = string("op_1090_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_1090_cast_fp16 = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, x = cache5_1_cast_fp16)[name = string("op_1090_cast_fp16")];
            bool var_1096_interleave_0 = const()[name = string("op_1096_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_1096_cast_fp16 = concat(axis = var_64, interleave = var_1096_interleave_0, values = (var_1090_cast_fp16, key_8_cast_fp16))[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85869824))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86918464))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_8_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_1100 = const()[name = string("op_1100"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_8_cast_fp16 = reshape(shape = var_1100, x = linear_30_cast_fp16)[name = string("q_8_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86919040))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87967680))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [4]> var_1104 = const()[name = string("op_1104"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_8_cast_fp16 = reshape(shape = var_1104, x = linear_31_cast_fp16)[name = string("k_8_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87968256))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89016896))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<int32, [4]> var_1108 = const()[name = string("op_1108"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_8_cast_fp16 = reshape(shape = var_1108, x = linear_32_cast_fp16)[name = string("v_8_cast_fp16")];
            tensor<int32, [4]> value_10_perm_0 = const()[name = string("value_10_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89017472)))];
            tensor<fp16, [1, 4, 8, 128]> var_1120_cast_fp16 = add(x = q_8_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89019584)))];
            tensor<fp16, [1, 4, 8, 128]> var_1122_cast_fp16 = add(x = q_8_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1122_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_8_perm_0 = const()[name = string("q_with_bias_v_8_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_71_transpose_x_0 = const()[name = string("x_71_transpose_x_0"), val = bool(false)];
            bool x_71_transpose_y_0 = const()[name = string("x_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_1124_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89021696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89143616))))[name = string("op_1124_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_8_cast_fp16 = transpose(perm = q_with_bias_v_8_perm_0, x = var_1122_cast_fp16)[name = string("transpose_332")];
            tensor<fp16, [1, 8, 4, 119]> x_71_cast_fp16 = matmul(transpose_x = x_71_transpose_x_0, transpose_y = x_71_transpose_y_0, x = q_with_bias_v_8_cast_fp16, y = op_1124_to_fp16_palettized)[name = string("x_71_cast_fp16")];
            tensor<int32, [8]> x0_10_pad_0 = const()[name = string("x0_10_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_10_mode_0 = const()[name = string("x0_10_mode_0"), val = string("constant")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_10_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x0_10_mode_0, pad = x0_10_pad_0, x = x_71_cast_fp16)[name = string("x0_10_cast_fp16")];
            tensor<int32, [4]> var_1132 = const()[name = string("op_1132"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_8_cast_fp16 = reshape(shape = var_1132, x = x0_10_cast_fp16)[name = string("x1_8_cast_fp16")];
            tensor<int32, [4]> var_1136_begin_0 = const()[name = string("op_1136_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1136_end_0 = const()[name = string("op_1136_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_1136_end_mask_0 = const()[name = string("op_1136_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, x = x1_8_cast_fp16)[name = string("op_1136_cast_fp16")];
            tensor<int32, [4]> var_1137 = const()[name = string("op_1137"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_8_cast_fp16 = reshape(shape = var_1137, x = var_1136_cast_fp16)[name = string("matrix_bd_8_cast_fp16")];
            bool matrix_ac_8_transpose_x_0 = const()[name = string("matrix_ac_8_transpose_x_0"), val = bool(false)];
            bool matrix_ac_8_transpose_y_0 = const()[name = string("matrix_ac_8_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_8_cast_fp16)[name = string("transpose_330")];
            tensor<fp16, [1, 8, 4, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1120_cast_fp16)[name = string("transpose_331")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_8_cast_fp16 = matmul(transpose_x = matrix_ac_8_transpose_x_0, transpose_y = matrix_ac_8_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_8_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_8_begin_0 = const()[name = string("matrix_bd0_8_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_8_end_0 = const()[name = string("matrix_bd0_8_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_8_end_mask_0 = const()[name = string("matrix_bd0_8_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_8_cast_fp16 = slice_by_index(begin = matrix_bd0_8_begin_0, end = matrix_bd0_8_end_0, end_mask = matrix_bd0_8_end_mask_0, x = matrix_bd_8_cast_fp16)[name = string("matrix_bd0_8_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1146_cast_fp16 = add(x = matrix_ac_8_cast_fp16, y = matrix_bd0_8_cast_fp16)[name = string("op_1146_cast_fp16")];
            fp16 _inversed_scores_8_y_0_to_fp16 = const()[name = string("_inversed_scores_8_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_8_cast_fp16 = mul(x = var_1146_cast_fp16, y = _inversed_scores_8_y_0_to_fp16)[name = string("_inversed_scores_8_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_8_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_8_cast_fp16, cond = mask0_4)[name = string("scores0_8_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1152_cast_fp16 = softmax(axis = var_56, x = scores0_8_cast_fp16)[name = string("op_1152_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_47_cast_fp16 = select(a = var_30_to_fp16, b = var_1152_cast_fp16, cond = mask0_4)[name = string("input0_47_cast_fp16")];
            bool x2_8_transpose_x_0 = const()[name = string("x2_8_transpose_x_0"), val = bool(false)];
            bool x2_8_transpose_y_0 = const()[name = string("x2_8_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_10_cast_fp16 = transpose(perm = value_10_perm_0, x = v_8_cast_fp16)[name = string("transpose_329")];
            tensor<fp16, [1, 8, 4, 128]> x2_8_cast_fp16 = matmul(transpose_x = x2_8_transpose_x_0, transpose_y = x2_8_transpose_y_0, x = input0_47_cast_fp16, y = value_10_cast_fp16)[name = string("x2_8_cast_fp16")];
            tensor<int32, [4]> var_1156_perm_0 = const()[name = string("op_1156_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1157 = const()[name = string("op_1157"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_1156_cast_fp16 = transpose(perm = var_1156_perm_0, x = x2_8_cast_fp16)[name = string("transpose_328")];
            tensor<fp16, [1, 4, 1024]> input1_24_cast_fp16 = reshape(shape = var_1157, x = var_1156_cast_fp16)[name = string("input1_24_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89144192))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90192832))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input1_24_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_49_cast_fp16 = add(x = input_49_cast_fp16, y = linear_34_cast_fp16)[name = string("input0_49_cast_fp16")];
            tensor<int32, [1]> x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90193408)))];
            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90195520)))];
            tensor<fp16, [1, 4, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input0_49_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<int32, [3]> input_53_perm_0 = const()[name = string("input_53_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_51_pad_type_0 = const()[name = string("input0_51_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_51_strides_0 = const()[name = string("input0_51_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_51_pad_0 = const()[name = string("input0_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_51_dilations_0 = const()[name = string("input0_51_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_51_groups_0 = const()[name = string("input0_51_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90197632))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92294848))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_53_cast_fp16 = transpose(perm = input_53_perm_0, x = x_75_cast_fp16)[name = string("transpose_327")];
            tensor<fp16, [1, 2048, 4]> input0_51_cast_fp16 = conv(dilations = input0_51_dilations_0, groups = input0_51_groups_0, pad = input0_51_pad_0, pad_type = input0_51_pad_type_0, strides = input0_51_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = string("input0_51_cast_fp16")];
            int32 x_77_split_num_splits_0 = const()[name = string("x_77_split_num_splits_0"), val = int32(2)];
            int32 x_77_split_axis_0 = const()[name = string("x_77_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_77_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_77_split_cast_fp16_1 = split(axis = x_77_split_axis_0, num_splits = x_77_split_num_splits_0, x = input0_51_cast_fp16)[name = string("x_77_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_77_split_1_sigmoid_cast_fp16 = sigmoid(x = x_77_split_cast_fp16_1)[name = string("x_77_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_77_cast_fp16 = mul(x = x_77_split_cast_fp16_0, y = x_77_split_1_sigmoid_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_53_cast_fp16 = select(a = var_30_to_fp16, b = x_77_cast_fp16, cond = var_570)[name = string("input0_53_cast_fp16")];
            bool new_x0_8_interleave_0 = const()[name = string("new_x0_8_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_8_cast_fp16 = concat(axis = var_56, interleave = new_x0_8_interleave_0, values = (cache6_1_cast_fp16, input0_53_cast_fp16))[name = string("new_x0_8_cast_fp16")];
            tensor<int32, [3]> var_1195_begin_0 = const()[name = string("op_1195_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1195_end_0 = const()[name = string("op_1195_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_1195_end_mask_0 = const()[name = string("op_1195_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_1195_cast_fp16 = slice_by_index(begin = var_1195_begin_0, end = var_1195_end_0, end_mask = var_1195_end_mask_0, x = new_x0_8_cast_fp16)[name = string("op_1195_cast_fp16")];
            string x_79_pad_type_0 = const()[name = string("x_79_pad_type_0"), val = string("valid")];
            int32 x_79_groups_0 = const()[name = string("x_79_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_79_strides_0 = const()[name = string("x_79_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_79_pad_0 = const()[name = string("x_79_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_79_dilations_0 = const()[name = string("x_79_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92295424))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92304704))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_79_cast_fp16 = conv(dilations = x_79_dilations_0, groups = x_79_groups_0, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = x_79_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_8_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> input1_26_perm_0 = const()[name = string("input1_26_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_81_axes_0 = const()[name = string("x_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92305280)))];
            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92307392)))];
            tensor<fp16, [1, 4, 1024]> input1_26_cast_fp16 = transpose(perm = input1_26_perm_0, x = x_79_cast_fp16)[name = string("transpose_326")];
            tensor<fp16, [1, 4, 1024]> x_81_cast_fp16 = layer_norm(axes = x_81_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input1_26_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [3]> input2_16_perm_0 = const()[name = string("input2_16_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_16_cast_fp16 = transpose(perm = input2_16_perm_0, x = x_81_cast_fp16)[name = string("transpose_325")];
            tensor<fp16, [1, 1024, 4]> var_1210_cast_fp16 = silu(x = input2_16_cast_fp16)[name = string("op_1210_cast_fp16")];
            string x_83_pad_type_0 = const()[name = string("x_83_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_83_strides_0 = const()[name = string("x_83_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_83_pad_0 = const()[name = string("x_83_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_83_dilations_0 = const()[name = string("x_83_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_83_groups_0 = const()[name = string("x_83_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92309504))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93358144))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_83_cast_fp16 = conv(dilations = x_83_dilations_0, groups = x_83_groups_0, pad = x_83_pad_0, pad_type = x_83_pad_type_0, strides = x_83_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_1210_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<int32, [3]> input3_10_perm_0 = const()[name = string("input3_10_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_10_cast_fp16 = transpose(perm = input3_10_perm_0, x = x_83_cast_fp16)[name = string("transpose_324")];
            tensor<fp16, [1, 4, 1024]> input1_28_cast_fp16 = add(x = input0_49_cast_fp16, y = input3_10_cast_fp16)[name = string("input1_28_cast_fp16")];
            tensor<int32, [1]> input0_55_axes_0 = const()[name = string("input0_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93358720)))];
            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93360832)))];
            tensor<fp16, [1, 4, 1024]> input0_55_cast_fp16 = layer_norm(axes = input0_55_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input1_28_cast_fp16)[name = string("input0_55_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93362944))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97557312))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_55_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1231_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("op_1231_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97557888))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101752256))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = var_1231_cast_fp16)[name = string("linear_36_cast_fp16")];
            fp16 var_1236_to_fp16 = const()[name = string("op_1236_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1237_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1236_to_fp16)[name = string("op_1237_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_18_cast_fp16 = add(x = input1_28_cast_fp16, y = var_1237_cast_fp16)[name = string("input2_18_cast_fp16")];
            tensor<int32, [1]> input0_57_axes_0 = const()[name = string("input0_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101752832)))];
            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101754944)))];
            tensor<fp16, [1, 4, 1024]> input0_57_cast_fp16 = layer_norm(axes = input0_57_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input2_18_cast_fp16)[name = string("input0_57_cast_fp16")];
            tensor<int32, [4]> cache7_1_begin_0 = const()[name = string("cache7_1_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> cache7_1_end_0 = const()[name = string("cache7_1_end_0"), val = tensor<int32, [4]>([5, 1, 56, 1024])];
            tensor<bool, [4]> cache7_1_end_mask_0 = const()[name = string("cache7_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache7_1_squeeze_mask_0 = const()[name = string("cache7_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache7_1_cast_fp16 = slice_by_index(begin = cache7_1_begin_0, end = cache7_1_end_0, end_mask = cache7_1_end_mask_0, squeeze_mask = cache7_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache7_1_cast_fp16")];
            tensor<int32, [4]> cache8_1_begin_0 = const()[name = string("cache8_1_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> cache8_1_end_0 = const()[name = string("cache8_1_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
            tensor<bool, [4]> cache8_1_end_mask_0 = const()[name = string("cache8_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache8_1_squeeze_mask_0 = const()[name = string("cache8_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache8_1_cast_fp16 = slice_by_index(begin = cache8_1_begin_0, end = cache8_1_end_0, end_mask = cache8_1_end_mask_0, squeeze_mask = cache8_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache8_1_cast_fp16")];
            tensor<int32, [1]> input_57_axes_0 = const()[name = string("input_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101757056)))];
            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101759168)))];
            tensor<fp16, [1, 4, 1024]> input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input0_57_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101761280))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105955648))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1266_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("op_1266_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105956224))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110150592))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = var_1266_cast_fp16)[name = string("linear_38_cast_fp16")];
            fp16 var_1271_to_fp16 = const()[name = string("op_1271_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1272_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1271_to_fp16)[name = string("op_1272_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_61_cast_fp16 = add(x = input0_57_cast_fp16, y = var_1272_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<int32, [1]> key_10_axes_0 = const()[name = string("key_10_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110151168)))];
            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110153280)))];
            tensor<fp16, [1, 4, 1024]> key_10_cast_fp16 = layer_norm(axes = key_10_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_61_cast_fp16)[name = string("key_10_cast_fp16")];
            bool input_63_interleave_0 = const()[name = string("input_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_63_cast_fp16 = concat(axis = var_64, interleave = input_63_interleave_0, values = (cache7_1_cast_fp16, key_10_cast_fp16))[name = string("input_63_cast_fp16")];
            tensor<int32, [3]> var_1294_begin_0 = const()[name = string("op_1294_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1294_end_0 = const()[name = string("op_1294_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_1294_end_mask_0 = const()[name = string("op_1294_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_1294_cast_fp16 = slice_by_index(begin = var_1294_begin_0, end = var_1294_end_0, end_mask = var_1294_end_mask_0, x = cache7_1_cast_fp16)[name = string("op_1294_cast_fp16")];
            bool var_1300_interleave_0 = const()[name = string("op_1300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_1300_cast_fp16 = concat(axis = var_64, interleave = var_1300_interleave_0, values = (var_1294_cast_fp16, key_10_cast_fp16))[name = string("op_1300_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110155392))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111204032))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_10_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_1304 = const()[name = string("op_1304"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_10_cast_fp16 = reshape(shape = var_1304, x = linear_39_cast_fp16)[name = string("q_10_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111204608))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112253248))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<int32, [4]> var_1308 = const()[name = string("op_1308"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_10_cast_fp16 = reshape(shape = var_1308, x = linear_40_cast_fp16)[name = string("k_10_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112253824))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113302464))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<int32, [4]> var_1312 = const()[name = string("op_1312"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_10_cast_fp16 = reshape(shape = var_1312, x = linear_41_cast_fp16)[name = string("v_10_cast_fp16")];
            tensor<int32, [4]> value_12_perm_0 = const()[name = string("value_12_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303040)))];
            tensor<fp16, [1, 4, 8, 128]> var_1324_cast_fp16 = add(x = q_10_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1324_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113305152)))];
            tensor<fp16, [1, 4, 8, 128]> var_1326_cast_fp16 = add(x = q_10_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_10_perm_0 = const()[name = string("q_with_bias_v_10_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_1328_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113307264))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113429184))))[name = string("op_1328_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_10_cast_fp16 = transpose(perm = q_with_bias_v_10_perm_0, x = var_1326_cast_fp16)[name = string("transpose_323")];
            tensor<fp16, [1, 8, 4, 119]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = q_with_bias_v_10_cast_fp16, y = op_1328_to_fp16_palettized)[name = string("x_91_cast_fp16")];
            tensor<int32, [8]> x0_12_pad_0 = const()[name = string("x0_12_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_12_mode_0 = const()[name = string("x0_12_mode_0"), val = string("constant")];
            fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_12_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x0_12_mode_0, pad = x0_12_pad_0, x = x_91_cast_fp16)[name = string("x0_12_cast_fp16")];
            tensor<int32, [4]> var_1336 = const()[name = string("op_1336"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_10_cast_fp16 = reshape(shape = var_1336, x = x0_12_cast_fp16)[name = string("x1_10_cast_fp16")];
            tensor<int32, [4]> var_1340_begin_0 = const()[name = string("op_1340_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1340_end_0 = const()[name = string("op_1340_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_1340_end_mask_0 = const()[name = string("op_1340_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, x = x1_10_cast_fp16)[name = string("op_1340_cast_fp16")];
            tensor<int32, [4]> var_1341 = const()[name = string("op_1341"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_10_cast_fp16 = reshape(shape = var_1341, x = var_1340_cast_fp16)[name = string("matrix_bd_10_cast_fp16")];
            bool matrix_ac_10_transpose_x_0 = const()[name = string("matrix_ac_10_transpose_x_0"), val = bool(false)];
            bool matrix_ac_10_transpose_y_0 = const()[name = string("matrix_ac_10_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_10_cast_fp16)[name = string("transpose_321")];
            tensor<fp16, [1, 8, 4, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1324_cast_fp16)[name = string("transpose_322")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_10_cast_fp16 = matmul(transpose_x = matrix_ac_10_transpose_x_0, transpose_y = matrix_ac_10_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_10_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_10_begin_0 = const()[name = string("matrix_bd0_10_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_10_end_0 = const()[name = string("matrix_bd0_10_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_10_end_mask_0 = const()[name = string("matrix_bd0_10_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_10_cast_fp16 = slice_by_index(begin = matrix_bd0_10_begin_0, end = matrix_bd0_10_end_0, end_mask = matrix_bd0_10_end_mask_0, x = matrix_bd_10_cast_fp16)[name = string("matrix_bd0_10_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1350_cast_fp16 = add(x = matrix_ac_10_cast_fp16, y = matrix_bd0_10_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 _inversed_scores_10_y_0_to_fp16 = const()[name = string("_inversed_scores_10_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_10_cast_fp16 = mul(x = var_1350_cast_fp16, y = _inversed_scores_10_y_0_to_fp16)[name = string("_inversed_scores_10_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_10_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_10_cast_fp16, cond = mask0_4)[name = string("scores0_10_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1356_cast_fp16 = softmax(axis = var_56, x = scores0_10_cast_fp16)[name = string("op_1356_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_59_cast_fp16 = select(a = var_30_to_fp16, b = var_1356_cast_fp16, cond = mask0_4)[name = string("input0_59_cast_fp16")];
            bool x2_10_transpose_x_0 = const()[name = string("x2_10_transpose_x_0"), val = bool(false)];
            bool x2_10_transpose_y_0 = const()[name = string("x2_10_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_12_cast_fp16 = transpose(perm = value_12_perm_0, x = v_10_cast_fp16)[name = string("transpose_320")];
            tensor<fp16, [1, 8, 4, 128]> x2_10_cast_fp16 = matmul(transpose_x = x2_10_transpose_x_0, transpose_y = x2_10_transpose_y_0, x = input0_59_cast_fp16, y = value_12_cast_fp16)[name = string("x2_10_cast_fp16")];
            tensor<int32, [4]> var_1360_perm_0 = const()[name = string("op_1360_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1361 = const()[name = string("op_1361"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_1360_cast_fp16 = transpose(perm = var_1360_perm_0, x = x2_10_cast_fp16)[name = string("transpose_319")];
            tensor<fp16, [1, 4, 1024]> input1_30_cast_fp16 = reshape(shape = var_1361, x = var_1360_cast_fp16)[name = string("input1_30_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113429760))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114478400))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input1_30_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_61_cast_fp16 = add(x = input_61_cast_fp16, y = linear_43_cast_fp16)[name = string("input0_61_cast_fp16")];
            tensor<int32, [1]> x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114478976)))];
            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114481088)))];
            tensor<fp16, [1, 4, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input0_61_cast_fp16)[name = string("x_95_cast_fp16")];
            tensor<int32, [3]> input_65_perm_0 = const()[name = string("input_65_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_63_pad_type_0 = const()[name = string("input0_63_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_63_strides_0 = const()[name = string("input0_63_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_63_pad_0 = const()[name = string("input0_63_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_63_dilations_0 = const()[name = string("input0_63_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_63_groups_0 = const()[name = string("input0_63_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114483200))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116580416))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_65_cast_fp16 = transpose(perm = input_65_perm_0, x = x_95_cast_fp16)[name = string("transpose_318")];
            tensor<fp16, [1, 2048, 4]> input0_63_cast_fp16 = conv(dilations = input0_63_dilations_0, groups = input0_63_groups_0, pad = input0_63_pad_0, pad_type = input0_63_pad_type_0, strides = input0_63_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = string("input0_63_cast_fp16")];
            int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)];
            int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input0_63_cast_fp16)[name = string("x_97_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_65_cast_fp16 = select(a = var_30_to_fp16, b = x_97_cast_fp16, cond = var_570)[name = string("input0_65_cast_fp16")];
            bool new_x0_10_interleave_0 = const()[name = string("new_x0_10_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_10_cast_fp16 = concat(axis = var_56, interleave = new_x0_10_interleave_0, values = (cache8_1_cast_fp16, input0_65_cast_fp16))[name = string("new_x0_10_cast_fp16")];
            tensor<int32, [3]> var_1399_begin_0 = const()[name = string("op_1399_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1399_end_0 = const()[name = string("op_1399_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_1399_end_mask_0 = const()[name = string("op_1399_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_1399_cast_fp16 = slice_by_index(begin = var_1399_begin_0, end = var_1399_end_0, end_mask = var_1399_end_mask_0, x = new_x0_10_cast_fp16)[name = string("op_1399_cast_fp16")];
            string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")];
            int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116580992))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116590272))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_10_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [3]> input1_32_perm_0 = const()[name = string("input1_32_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116590848)))];
            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116592960)))];
            tensor<fp16, [1, 4, 1024]> input1_32_cast_fp16 = transpose(perm = input1_32_perm_0, x = x_99_cast_fp16)[name = string("transpose_317")];
            tensor<fp16, [1, 4, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input1_32_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [3]> input2_20_perm_0 = const()[name = string("input2_20_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_20_cast_fp16 = transpose(perm = input2_20_perm_0, x = x_101_cast_fp16)[name = string("transpose_316")];
            tensor<fp16, [1, 1024, 4]> var_1414_cast_fp16 = silu(x = input2_20_cast_fp16)[name = string("op_1414_cast_fp16")];
            string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116595072))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117643712))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_1414_cast_fp16)[name = string("x_103_cast_fp16")];
            tensor<int32, [3]> input3_12_perm_0 = const()[name = string("input3_12_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_12_cast_fp16 = transpose(perm = input3_12_perm_0, x = x_103_cast_fp16)[name = string("transpose_315")];
            tensor<fp16, [1, 4, 1024]> input1_34_cast_fp16 = add(x = input0_61_cast_fp16, y = input3_12_cast_fp16)[name = string("input1_34_cast_fp16")];
            tensor<int32, [1]> input0_67_axes_0 = const()[name = string("input0_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117644288)))];
            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117646400)))];
            tensor<fp16, [1, 4, 1024]> input0_67_cast_fp16 = layer_norm(axes = input0_67_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input1_34_cast_fp16)[name = string("input0_67_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117648512))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121842880))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_67_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1435_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("op_1435_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121843456))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126037824))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = var_1435_cast_fp16)[name = string("linear_45_cast_fp16")];
            fp16 var_1440_to_fp16 = const()[name = string("op_1440_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1441_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1440_to_fp16)[name = string("op_1441_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_22_cast_fp16 = add(x = input1_34_cast_fp16, y = var_1441_cast_fp16)[name = string("input2_22_cast_fp16")];
            tensor<int32, [1]> input0_69_axes_0 = const()[name = string("input0_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126038400)))];
            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126040512)))];
            tensor<fp16, [1, 4, 1024]> input0_69_cast_fp16 = layer_norm(axes = input0_69_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input2_22_cast_fp16)[name = string("input0_69_cast_fp16")];
            tensor<int32, [4]> cache9_1_begin_0 = const()[name = string("cache9_1_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> cache9_1_end_0 = const()[name = string("cache9_1_end_0"), val = tensor<int32, [4]>([6, 1, 56, 1024])];
            tensor<bool, [4]> cache9_1_end_mask_0 = const()[name = string("cache9_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache9_1_squeeze_mask_0 = const()[name = string("cache9_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache9_1_cast_fp16 = slice_by_index(begin = cache9_1_begin_0, end = cache9_1_end_0, end_mask = cache9_1_end_mask_0, squeeze_mask = cache9_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache9_1_cast_fp16")];
            tensor<int32, [4]> cache10_1_begin_0 = const()[name = string("cache10_1_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> cache10_1_end_0 = const()[name = string("cache10_1_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
            tensor<bool, [4]> cache10_1_end_mask_0 = const()[name = string("cache10_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache10_1_squeeze_mask_0 = const()[name = string("cache10_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache10_1_cast_fp16 = slice_by_index(begin = cache10_1_begin_0, end = cache10_1_end_0, end_mask = cache10_1_end_mask_0, squeeze_mask = cache10_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache10_1_cast_fp16")];
            tensor<int32, [1]> input_69_axes_0 = const()[name = string("input_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126042624)))];
            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126044736)))];
            tensor<fp16, [1, 4, 1024]> input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input0_69_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126046848))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130241216))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1470_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130241792))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134436160))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = var_1470_cast_fp16)[name = string("linear_47_cast_fp16")];
            fp16 var_1475_to_fp16 = const()[name = string("op_1475_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1476_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1475_to_fp16)[name = string("op_1476_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_73_cast_fp16 = add(x = input0_69_cast_fp16, y = var_1476_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<int32, [1]> key_12_axes_0 = const()[name = string("key_12_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134436736)))];
            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134438848)))];
            tensor<fp16, [1, 4, 1024]> key_12_cast_fp16 = layer_norm(axes = key_12_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_73_cast_fp16)[name = string("key_12_cast_fp16")];
            bool input_75_interleave_0 = const()[name = string("input_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_75_cast_fp16 = concat(axis = var_64, interleave = input_75_interleave_0, values = (cache9_1_cast_fp16, key_12_cast_fp16))[name = string("input_75_cast_fp16")];
            tensor<int32, [3]> var_1498_begin_0 = const()[name = string("op_1498_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1498_end_0 = const()[name = string("op_1498_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_1498_end_mask_0 = const()[name = string("op_1498_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_1498_cast_fp16 = slice_by_index(begin = var_1498_begin_0, end = var_1498_end_0, end_mask = var_1498_end_mask_0, x = cache9_1_cast_fp16)[name = string("op_1498_cast_fp16")];
            bool var_1504_interleave_0 = const()[name = string("op_1504_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_1504_cast_fp16 = concat(axis = var_64, interleave = var_1504_interleave_0, values = (var_1498_cast_fp16, key_12_cast_fp16))[name = string("op_1504_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134440960))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135489600))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_12_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_1508 = const()[name = string("op_1508"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_12_cast_fp16 = reshape(shape = var_1508, x = linear_48_cast_fp16)[name = string("q_12_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135490176))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136538816))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1512 = const()[name = string("op_1512"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_12_cast_fp16 = reshape(shape = var_1512, x = linear_49_cast_fp16)[name = string("k_12_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136539392))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137588032))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1516 = const()[name = string("op_1516"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_12_cast_fp16 = reshape(shape = var_1516, x = linear_50_cast_fp16)[name = string("v_12_cast_fp16")];
            tensor<int32, [4]> value_14_perm_0 = const()[name = string("value_14_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137588608)))];
            tensor<fp16, [1, 4, 8, 128]> var_1528_cast_fp16 = add(x = q_12_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1528_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137590720)))];
            tensor<fp16, [1, 4, 8, 128]> var_1530_cast_fp16 = add(x = q_12_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_12_perm_0 = const()[name = string("q_with_bias_v_12_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_1532_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137592832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137714752))))[name = string("op_1532_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_12_cast_fp16 = transpose(perm = q_with_bias_v_12_perm_0, x = var_1530_cast_fp16)[name = string("transpose_314")];
            tensor<fp16, [1, 8, 4, 119]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_12_cast_fp16, y = op_1532_to_fp16_palettized)[name = string("x_111_cast_fp16")];
            tensor<int32, [8]> x0_14_pad_0 = const()[name = string("x0_14_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_14_mode_0 = const()[name = string("x0_14_mode_0"), val = string("constant")];
            fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_14_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x0_14_mode_0, pad = x0_14_pad_0, x = x_111_cast_fp16)[name = string("x0_14_cast_fp16")];
            tensor<int32, [4]> var_1540 = const()[name = string("op_1540"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_12_cast_fp16 = reshape(shape = var_1540, x = x0_14_cast_fp16)[name = string("x1_12_cast_fp16")];
            tensor<int32, [4]> var_1544_begin_0 = const()[name = string("op_1544_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1544_end_0 = const()[name = string("op_1544_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_1544_end_mask_0 = const()[name = string("op_1544_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_1544_cast_fp16 = slice_by_index(begin = var_1544_begin_0, end = var_1544_end_0, end_mask = var_1544_end_mask_0, x = x1_12_cast_fp16)[name = string("op_1544_cast_fp16")];
            tensor<int32, [4]> var_1545 = const()[name = string("op_1545"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_12_cast_fp16 = reshape(shape = var_1545, x = var_1544_cast_fp16)[name = string("matrix_bd_12_cast_fp16")];
            bool matrix_ac_12_transpose_x_0 = const()[name = string("matrix_ac_12_transpose_x_0"), val = bool(false)];
            bool matrix_ac_12_transpose_y_0 = const()[name = string("matrix_ac_12_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_12_cast_fp16)[name = string("transpose_312")];
            tensor<fp16, [1, 8, 4, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1528_cast_fp16)[name = string("transpose_313")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_12_cast_fp16 = matmul(transpose_x = matrix_ac_12_transpose_x_0, transpose_y = matrix_ac_12_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_12_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_12_begin_0 = const()[name = string("matrix_bd0_12_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_12_end_0 = const()[name = string("matrix_bd0_12_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_12_end_mask_0 = const()[name = string("matrix_bd0_12_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_12_cast_fp16 = slice_by_index(begin = matrix_bd0_12_begin_0, end = matrix_bd0_12_end_0, end_mask = matrix_bd0_12_end_mask_0, x = matrix_bd_12_cast_fp16)[name = string("matrix_bd0_12_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1554_cast_fp16 = add(x = matrix_ac_12_cast_fp16, y = matrix_bd0_12_cast_fp16)[name = string("op_1554_cast_fp16")];
            fp16 _inversed_scores_12_y_0_to_fp16 = const()[name = string("_inversed_scores_12_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_12_cast_fp16 = mul(x = var_1554_cast_fp16, y = _inversed_scores_12_y_0_to_fp16)[name = string("_inversed_scores_12_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_12_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_12_cast_fp16, cond = mask0_4)[name = string("scores0_12_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1560_cast_fp16 = softmax(axis = var_56, x = scores0_12_cast_fp16)[name = string("op_1560_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_71_cast_fp16 = select(a = var_30_to_fp16, b = var_1560_cast_fp16, cond = mask0_4)[name = string("input0_71_cast_fp16")];
            bool x2_12_transpose_x_0 = const()[name = string("x2_12_transpose_x_0"), val = bool(false)];
            bool x2_12_transpose_y_0 = const()[name = string("x2_12_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_14_cast_fp16 = transpose(perm = value_14_perm_0, x = v_12_cast_fp16)[name = string("transpose_311")];
            tensor<fp16, [1, 8, 4, 128]> x2_12_cast_fp16 = matmul(transpose_x = x2_12_transpose_x_0, transpose_y = x2_12_transpose_y_0, x = input0_71_cast_fp16, y = value_14_cast_fp16)[name = string("x2_12_cast_fp16")];
            tensor<int32, [4]> var_1564_perm_0 = const()[name = string("op_1564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1565 = const()[name = string("op_1565"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_1564_cast_fp16 = transpose(perm = var_1564_perm_0, x = x2_12_cast_fp16)[name = string("transpose_310")];
            tensor<fp16, [1, 4, 1024]> input1_36_cast_fp16 = reshape(shape = var_1565, x = var_1564_cast_fp16)[name = string("input1_36_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137715328))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138763968))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input1_36_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_73_cast_fp16 = add(x = input_73_cast_fp16, y = linear_52_cast_fp16)[name = string("input0_73_cast_fp16")];
            tensor<int32, [1]> x_115_axes_0 = const()[name = string("x_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138764544)))];
            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138766656)))];
            tensor<fp16, [1, 4, 1024]> x_115_cast_fp16 = layer_norm(axes = x_115_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input0_73_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<int32, [3]> input_77_perm_0 = const()[name = string("input_77_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_75_pad_type_0 = const()[name = string("input0_75_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_75_strides_0 = const()[name = string("input0_75_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_75_pad_0 = const()[name = string("input0_75_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_75_dilations_0 = const()[name = string("input0_75_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_75_groups_0 = const()[name = string("input0_75_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138768768))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140865984))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_77_cast_fp16 = transpose(perm = input_77_perm_0, x = x_115_cast_fp16)[name = string("transpose_309")];
            tensor<fp16, [1, 2048, 4]> input0_75_cast_fp16 = conv(dilations = input0_75_dilations_0, groups = input0_75_groups_0, pad = input0_75_pad_0, pad_type = input0_75_pad_type_0, strides = input0_75_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = string("input0_75_cast_fp16")];
            int32 x_117_split_num_splits_0 = const()[name = string("x_117_split_num_splits_0"), val = int32(2)];
            int32 x_117_split_axis_0 = const()[name = string("x_117_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_117_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_117_split_cast_fp16_1 = split(axis = x_117_split_axis_0, num_splits = x_117_split_num_splits_0, x = input0_75_cast_fp16)[name = string("x_117_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_117_split_1_sigmoid_cast_fp16 = sigmoid(x = x_117_split_cast_fp16_1)[name = string("x_117_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_117_cast_fp16 = mul(x = x_117_split_cast_fp16_0, y = x_117_split_1_sigmoid_cast_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_77_cast_fp16 = select(a = var_30_to_fp16, b = x_117_cast_fp16, cond = var_570)[name = string("input0_77_cast_fp16")];
            bool new_x0_12_interleave_0 = const()[name = string("new_x0_12_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_12_cast_fp16 = concat(axis = var_56, interleave = new_x0_12_interleave_0, values = (cache10_1_cast_fp16, input0_77_cast_fp16))[name = string("new_x0_12_cast_fp16")];
            tensor<int32, [3]> var_1603_begin_0 = const()[name = string("op_1603_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1603_end_0 = const()[name = string("op_1603_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_1603_end_mask_0 = const()[name = string("op_1603_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_1603_cast_fp16 = slice_by_index(begin = var_1603_begin_0, end = var_1603_end_0, end_mask = var_1603_end_mask_0, x = new_x0_12_cast_fp16)[name = string("op_1603_cast_fp16")];
            string x_119_pad_type_0 = const()[name = string("x_119_pad_type_0"), val = string("valid")];
            int32 x_119_groups_0 = const()[name = string("x_119_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_119_strides_0 = const()[name = string("x_119_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_119_pad_0 = const()[name = string("x_119_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_119_dilations_0 = const()[name = string("x_119_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140866560))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140875840))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_119_cast_fp16 = conv(dilations = x_119_dilations_0, groups = x_119_groups_0, pad = x_119_pad_0, pad_type = x_119_pad_type_0, strides = x_119_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_12_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<int32, [3]> input1_38_perm_0 = const()[name = string("input1_38_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140876416)))];
            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140878528)))];
            tensor<fp16, [1, 4, 1024]> input1_38_cast_fp16 = transpose(perm = input1_38_perm_0, x = x_119_cast_fp16)[name = string("transpose_308")];
            tensor<fp16, [1, 4, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input1_38_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [3]> input2_24_perm_0 = const()[name = string("input2_24_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_24_cast_fp16 = transpose(perm = input2_24_perm_0, x = x_121_cast_fp16)[name = string("transpose_307")];
            tensor<fp16, [1, 1024, 4]> var_1618_cast_fp16 = silu(x = input2_24_cast_fp16)[name = string("op_1618_cast_fp16")];
            string x_123_pad_type_0 = const()[name = string("x_123_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_123_strides_0 = const()[name = string("x_123_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_123_pad_0 = const()[name = string("x_123_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_123_dilations_0 = const()[name = string("x_123_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_123_groups_0 = const()[name = string("x_123_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140880640))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141929280))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_123_cast_fp16 = conv(dilations = x_123_dilations_0, groups = x_123_groups_0, pad = x_123_pad_0, pad_type = x_123_pad_type_0, strides = x_123_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_1618_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [3]> input3_14_perm_0 = const()[name = string("input3_14_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_14_cast_fp16 = transpose(perm = input3_14_perm_0, x = x_123_cast_fp16)[name = string("transpose_306")];
            tensor<fp16, [1, 4, 1024]> input1_40_cast_fp16 = add(x = input0_73_cast_fp16, y = input3_14_cast_fp16)[name = string("input1_40_cast_fp16")];
            tensor<int32, [1]> input0_79_axes_0 = const()[name = string("input0_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141929856)))];
            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141931968)))];
            tensor<fp16, [1, 4, 1024]> input0_79_cast_fp16 = layer_norm(axes = input0_79_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input1_40_cast_fp16)[name = string("input0_79_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141934080))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146128448))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_79_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1639_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1639_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146129024))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150323392))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = var_1639_cast_fp16)[name = string("linear_54_cast_fp16")];
            fp16 var_1644_to_fp16 = const()[name = string("op_1644_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1645_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1644_to_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_26_cast_fp16 = add(x = input1_40_cast_fp16, y = var_1645_cast_fp16)[name = string("input2_26_cast_fp16")];
            tensor<int32, [1]> input0_81_axes_0 = const()[name = string("input0_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150323968)))];
            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150326080)))];
            tensor<fp16, [1, 4, 1024]> input0_81_cast_fp16 = layer_norm(axes = input0_81_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input2_26_cast_fp16)[name = string("input0_81_cast_fp16")];
            tensor<int32, [4]> cache11_1_begin_0 = const()[name = string("cache11_1_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> cache11_1_end_0 = const()[name = string("cache11_1_end_0"), val = tensor<int32, [4]>([7, 1, 56, 1024])];
            tensor<bool, [4]> cache11_1_end_mask_0 = const()[name = string("cache11_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache11_1_squeeze_mask_0 = const()[name = string("cache11_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache11_1_cast_fp16 = slice_by_index(begin = cache11_1_begin_0, end = cache11_1_end_0, end_mask = cache11_1_end_mask_0, squeeze_mask = cache11_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache11_1_cast_fp16")];
            tensor<int32, [4]> cache12_1_begin_0 = const()[name = string("cache12_1_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> cache12_1_end_0 = const()[name = string("cache12_1_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
            tensor<bool, [4]> cache12_1_end_mask_0 = const()[name = string("cache12_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache12_1_squeeze_mask_0 = const()[name = string("cache12_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache12_1_cast_fp16 = slice_by_index(begin = cache12_1_begin_0, end = cache12_1_end_0, end_mask = cache12_1_end_mask_0, squeeze_mask = cache12_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache12_1_cast_fp16")];
            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150328192)))];
            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150330304)))];
            tensor<fp16, [1, 4, 1024]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input0_81_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150332416))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154526784))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1674_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("op_1674_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154527360))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158721728))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = var_1674_cast_fp16)[name = string("linear_56_cast_fp16")];
            fp16 var_1679_to_fp16 = const()[name = string("op_1679_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1680_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1679_to_fp16)[name = string("op_1680_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_85_cast_fp16 = add(x = input0_81_cast_fp16, y = var_1680_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<int32, [1]> key_14_axes_0 = const()[name = string("key_14_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158722304)))];
            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158724416)))];
            tensor<fp16, [1, 4, 1024]> key_14_cast_fp16 = layer_norm(axes = key_14_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_85_cast_fp16)[name = string("key_14_cast_fp16")];
            bool input_87_interleave_0 = const()[name = string("input_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_87_cast_fp16 = concat(axis = var_64, interleave = input_87_interleave_0, values = (cache11_1_cast_fp16, key_14_cast_fp16))[name = string("input_87_cast_fp16")];
            tensor<int32, [3]> var_1702_begin_0 = const()[name = string("op_1702_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1702_end_0 = const()[name = string("op_1702_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_1702_end_mask_0 = const()[name = string("op_1702_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, x = cache11_1_cast_fp16)[name = string("op_1702_cast_fp16")];
            bool var_1708_interleave_0 = const()[name = string("op_1708_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_1708_cast_fp16 = concat(axis = var_64, interleave = var_1708_interleave_0, values = (var_1702_cast_fp16, key_14_cast_fp16))[name = string("op_1708_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158726528))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159775168))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_14_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1712 = const()[name = string("op_1712"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_14_cast_fp16 = reshape(shape = var_1712, x = linear_57_cast_fp16)[name = string("q_14_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159775744))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160824384))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1716 = const()[name = string("op_1716"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_14_cast_fp16 = reshape(shape = var_1716, x = linear_58_cast_fp16)[name = string("k_14_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160824960))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161873600))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<int32, [4]> var_1720 = const()[name = string("op_1720"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_14_cast_fp16 = reshape(shape = var_1720, x = linear_59_cast_fp16)[name = string("v_14_cast_fp16")];
            tensor<int32, [4]> value_16_perm_0 = const()[name = string("value_16_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161874176)))];
            tensor<fp16, [1, 4, 8, 128]> var_1732_cast_fp16 = add(x = q_14_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1732_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161876288)))];
            tensor<fp16, [1, 4, 8, 128]> var_1734_cast_fp16 = add(x = q_14_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1734_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_14_perm_0 = const()[name = string("q_with_bias_v_14_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_131_transpose_x_0 = const()[name = string("x_131_transpose_x_0"), val = bool(false)];
            bool x_131_transpose_y_0 = const()[name = string("x_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_1736_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161878400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162000320))))[name = string("op_1736_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_14_cast_fp16 = transpose(perm = q_with_bias_v_14_perm_0, x = var_1734_cast_fp16)[name = string("transpose_305")];
            tensor<fp16, [1, 8, 4, 119]> x_131_cast_fp16 = matmul(transpose_x = x_131_transpose_x_0, transpose_y = x_131_transpose_y_0, x = q_with_bias_v_14_cast_fp16, y = op_1736_to_fp16_palettized)[name = string("x_131_cast_fp16")];
            tensor<int32, [8]> x0_16_pad_0 = const()[name = string("x0_16_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_16_mode_0 = const()[name = string("x0_16_mode_0"), val = string("constant")];
            fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_16_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x0_16_mode_0, pad = x0_16_pad_0, x = x_131_cast_fp16)[name = string("x0_16_cast_fp16")];
            tensor<int32, [4]> var_1744 = const()[name = string("op_1744"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_14_cast_fp16 = reshape(shape = var_1744, x = x0_16_cast_fp16)[name = string("x1_14_cast_fp16")];
            tensor<int32, [4]> var_1748_begin_0 = const()[name = string("op_1748_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1748_end_0 = const()[name = string("op_1748_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_1748_end_mask_0 = const()[name = string("op_1748_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_1748_cast_fp16 = slice_by_index(begin = var_1748_begin_0, end = var_1748_end_0, end_mask = var_1748_end_mask_0, x = x1_14_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_14_cast_fp16 = reshape(shape = var_1749, x = var_1748_cast_fp16)[name = string("matrix_bd_14_cast_fp16")];
            bool matrix_ac_14_transpose_x_0 = const()[name = string("matrix_ac_14_transpose_x_0"), val = bool(false)];
            bool matrix_ac_14_transpose_y_0 = const()[name = string("matrix_ac_14_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_14_cast_fp16)[name = string("transpose_303")];
            tensor<fp16, [1, 8, 4, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1732_cast_fp16)[name = string("transpose_304")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_14_cast_fp16 = matmul(transpose_x = matrix_ac_14_transpose_x_0, transpose_y = matrix_ac_14_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_14_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_14_begin_0 = const()[name = string("matrix_bd0_14_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_14_end_0 = const()[name = string("matrix_bd0_14_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_14_end_mask_0 = const()[name = string("matrix_bd0_14_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_14_cast_fp16 = slice_by_index(begin = matrix_bd0_14_begin_0, end = matrix_bd0_14_end_0, end_mask = matrix_bd0_14_end_mask_0, x = matrix_bd_14_cast_fp16)[name = string("matrix_bd0_14_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1758_cast_fp16 = add(x = matrix_ac_14_cast_fp16, y = matrix_bd0_14_cast_fp16)[name = string("op_1758_cast_fp16")];
            fp16 _inversed_scores_14_y_0_to_fp16 = const()[name = string("_inversed_scores_14_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_14_cast_fp16 = mul(x = var_1758_cast_fp16, y = _inversed_scores_14_y_0_to_fp16)[name = string("_inversed_scores_14_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_14_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_14_cast_fp16, cond = mask0_4)[name = string("scores0_14_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1764_cast_fp16 = softmax(axis = var_56, x = scores0_14_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_83_cast_fp16 = select(a = var_30_to_fp16, b = var_1764_cast_fp16, cond = mask0_4)[name = string("input0_83_cast_fp16")];
            bool x2_14_transpose_x_0 = const()[name = string("x2_14_transpose_x_0"), val = bool(false)];
            bool x2_14_transpose_y_0 = const()[name = string("x2_14_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_16_cast_fp16 = transpose(perm = value_16_perm_0, x = v_14_cast_fp16)[name = string("transpose_302")];
            tensor<fp16, [1, 8, 4, 128]> x2_14_cast_fp16 = matmul(transpose_x = x2_14_transpose_x_0, transpose_y = x2_14_transpose_y_0, x = input0_83_cast_fp16, y = value_16_cast_fp16)[name = string("x2_14_cast_fp16")];
            tensor<int32, [4]> var_1768_perm_0 = const()[name = string("op_1768_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1769 = const()[name = string("op_1769"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_1768_cast_fp16 = transpose(perm = var_1768_perm_0, x = x2_14_cast_fp16)[name = string("transpose_301")];
            tensor<fp16, [1, 4, 1024]> input1_42_cast_fp16 = reshape(shape = var_1769, x = var_1768_cast_fp16)[name = string("input1_42_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162000896))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163049536))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input1_42_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_85_cast_fp16 = add(x = input_85_cast_fp16, y = linear_61_cast_fp16)[name = string("input0_85_cast_fp16")];
            tensor<int32, [1]> x_135_axes_0 = const()[name = string("x_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163050112)))];
            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163052224)))];
            tensor<fp16, [1, 4, 1024]> x_135_cast_fp16 = layer_norm(axes = x_135_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input0_85_cast_fp16)[name = string("x_135_cast_fp16")];
            tensor<int32, [3]> input_89_perm_0 = const()[name = string("input_89_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_87_pad_type_0 = const()[name = string("input0_87_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_87_strides_0 = const()[name = string("input0_87_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_87_pad_0 = const()[name = string("input0_87_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_87_dilations_0 = const()[name = string("input0_87_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_87_groups_0 = const()[name = string("input0_87_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163054336))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165151552))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_89_cast_fp16 = transpose(perm = input_89_perm_0, x = x_135_cast_fp16)[name = string("transpose_300")];
            tensor<fp16, [1, 2048, 4]> input0_87_cast_fp16 = conv(dilations = input0_87_dilations_0, groups = input0_87_groups_0, pad = input0_87_pad_0, pad_type = input0_87_pad_type_0, strides = input0_87_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = string("input0_87_cast_fp16")];
            int32 x_137_split_num_splits_0 = const()[name = string("x_137_split_num_splits_0"), val = int32(2)];
            int32 x_137_split_axis_0 = const()[name = string("x_137_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_137_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_137_split_cast_fp16_1 = split(axis = x_137_split_axis_0, num_splits = x_137_split_num_splits_0, x = input0_87_cast_fp16)[name = string("x_137_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_137_split_1_sigmoid_cast_fp16 = sigmoid(x = x_137_split_cast_fp16_1)[name = string("x_137_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_137_cast_fp16 = mul(x = x_137_split_cast_fp16_0, y = x_137_split_1_sigmoid_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_89_cast_fp16 = select(a = var_30_to_fp16, b = x_137_cast_fp16, cond = var_570)[name = string("input0_89_cast_fp16")];
            bool new_x0_14_interleave_0 = const()[name = string("new_x0_14_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_14_cast_fp16 = concat(axis = var_56, interleave = new_x0_14_interleave_0, values = (cache12_1_cast_fp16, input0_89_cast_fp16))[name = string("new_x0_14_cast_fp16")];
            tensor<int32, [3]> var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = new_x0_14_cast_fp16)[name = string("op_1807_cast_fp16")];
            string x_139_pad_type_0 = const()[name = string("x_139_pad_type_0"), val = string("valid")];
            int32 x_139_groups_0 = const()[name = string("x_139_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_139_strides_0 = const()[name = string("x_139_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_139_dilations_0 = const()[name = string("x_139_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165152128))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165161408))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_139_cast_fp16 = conv(dilations = x_139_dilations_0, groups = x_139_groups_0, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = x_139_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_14_cast_fp16)[name = string("x_139_cast_fp16")];
            tensor<int32, [3]> input1_44_perm_0 = const()[name = string("input1_44_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_141_axes_0 = const()[name = string("x_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165161984)))];
            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165164096)))];
            tensor<fp16, [1, 4, 1024]> input1_44_cast_fp16 = transpose(perm = input1_44_perm_0, x = x_139_cast_fp16)[name = string("transpose_299")];
            tensor<fp16, [1, 4, 1024]> x_141_cast_fp16 = layer_norm(axes = x_141_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input1_44_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [3]> input2_28_perm_0 = const()[name = string("input2_28_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_28_cast_fp16 = transpose(perm = input2_28_perm_0, x = x_141_cast_fp16)[name = string("transpose_298")];
            tensor<fp16, [1, 1024, 4]> var_1822_cast_fp16 = silu(x = input2_28_cast_fp16)[name = string("op_1822_cast_fp16")];
            string x_143_pad_type_0 = const()[name = string("x_143_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_143_strides_0 = const()[name = string("x_143_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_143_pad_0 = const()[name = string("x_143_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_143_dilations_0 = const()[name = string("x_143_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_143_groups_0 = const()[name = string("x_143_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165166208))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166214848))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_143_cast_fp16 = conv(dilations = x_143_dilations_0, groups = x_143_groups_0, pad = x_143_pad_0, pad_type = x_143_pad_type_0, strides = x_143_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_1822_cast_fp16)[name = string("x_143_cast_fp16")];
            tensor<int32, [3]> input3_16_perm_0 = const()[name = string("input3_16_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_16_cast_fp16 = transpose(perm = input3_16_perm_0, x = x_143_cast_fp16)[name = string("transpose_297")];
            tensor<fp16, [1, 4, 1024]> input1_46_cast_fp16 = add(x = input0_85_cast_fp16, y = input3_16_cast_fp16)[name = string("input1_46_cast_fp16")];
            tensor<int32, [1]> input0_91_axes_0 = const()[name = string("input0_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166215424)))];
            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166217536)))];
            tensor<fp16, [1, 4, 1024]> input0_91_cast_fp16 = layer_norm(axes = input0_91_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input1_46_cast_fp16)[name = string("input0_91_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166219648))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170414016))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_91_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1843_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("op_1843_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170414592))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174608960))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = var_1843_cast_fp16)[name = string("linear_63_cast_fp16")];
            fp16 var_1848_to_fp16 = const()[name = string("op_1848_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1849_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1848_to_fp16)[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_30_cast_fp16 = add(x = input1_46_cast_fp16, y = var_1849_cast_fp16)[name = string("input2_30_cast_fp16")];
            tensor<int32, [1]> input0_93_axes_0 = const()[name = string("input0_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174609536)))];
            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174611648)))];
            tensor<fp16, [1, 4, 1024]> input0_93_cast_fp16 = layer_norm(axes = input0_93_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input2_30_cast_fp16)[name = string("input0_93_cast_fp16")];
            tensor<int32, [4]> cache13_1_begin_0 = const()[name = string("cache13_1_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> cache13_1_end_0 = const()[name = string("cache13_1_end_0"), val = tensor<int32, [4]>([8, 1, 56, 1024])];
            tensor<bool, [4]> cache13_1_end_mask_0 = const()[name = string("cache13_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache13_1_squeeze_mask_0 = const()[name = string("cache13_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache13_1_cast_fp16 = slice_by_index(begin = cache13_1_begin_0, end = cache13_1_end_0, end_mask = cache13_1_end_mask_0, squeeze_mask = cache13_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache13_1_cast_fp16")];
            tensor<int32, [4]> cache14_1_begin_0 = const()[name = string("cache14_1_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> cache14_1_end_0 = const()[name = string("cache14_1_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
            tensor<bool, [4]> cache14_1_end_mask_0 = const()[name = string("cache14_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache14_1_squeeze_mask_0 = const()[name = string("cache14_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache14_1_cast_fp16 = slice_by_index(begin = cache14_1_begin_0, end = cache14_1_end_0, end_mask = cache14_1_end_mask_0, squeeze_mask = cache14_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache14_1_cast_fp16")];
            tensor<int32, [1]> input_93_axes_0 = const()[name = string("input_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174613760)))];
            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174615872)))];
            tensor<fp16, [1, 4, 1024]> input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input0_93_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174617984))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178812352))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_93_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_1878_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("op_1878_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178812928))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183007296))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = var_1878_cast_fp16)[name = string("linear_65_cast_fp16")];
            fp16 var_1883_to_fp16 = const()[name = string("op_1883_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_1884_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1883_to_fp16)[name = string("op_1884_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_97_cast_fp16 = add(x = input0_93_cast_fp16, y = var_1884_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<int32, [1]> key_16_axes_0 = const()[name = string("key_16_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183007872)))];
            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183009984)))];
            tensor<fp16, [1, 4, 1024]> key_16_cast_fp16 = layer_norm(axes = key_16_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_97_cast_fp16)[name = string("key_16_cast_fp16")];
            bool input_99_interleave_0 = const()[name = string("input_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_99_cast_fp16 = concat(axis = var_64, interleave = input_99_interleave_0, values = (cache13_1_cast_fp16, key_16_cast_fp16))[name = string("input_99_cast_fp16")];
            tensor<int32, [3]> var_1906_begin_0 = const()[name = string("op_1906_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1906_end_0 = const()[name = string("op_1906_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_1906_end_mask_0 = const()[name = string("op_1906_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_1906_cast_fp16 = slice_by_index(begin = var_1906_begin_0, end = var_1906_end_0, end_mask = var_1906_end_mask_0, x = cache13_1_cast_fp16)[name = string("op_1906_cast_fp16")];
            bool var_1912_interleave_0 = const()[name = string("op_1912_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_1912_cast_fp16 = concat(axis = var_64, interleave = var_1912_interleave_0, values = (var_1906_cast_fp16, key_16_cast_fp16))[name = string("op_1912_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183012096))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184060736))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_16_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_1916 = const()[name = string("op_1916"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_16_cast_fp16 = reshape(shape = var_1916, x = linear_66_cast_fp16)[name = string("q_16_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184061312))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185109952))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<int32, [4]> var_1920 = const()[name = string("op_1920"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_16_cast_fp16 = reshape(shape = var_1920, x = linear_67_cast_fp16)[name = string("k_16_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185110528))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186159168))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [4]> var_1924 = const()[name = string("op_1924"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_16_cast_fp16 = reshape(shape = var_1924, x = linear_68_cast_fp16)[name = string("v_16_cast_fp16")];
            tensor<int32, [4]> value_18_perm_0 = const()[name = string("value_18_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186159744)))];
            tensor<fp16, [1, 4, 8, 128]> var_1936_cast_fp16 = add(x = q_16_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186161856)))];
            tensor<fp16, [1, 4, 8, 128]> var_1938_cast_fp16 = add(x = q_16_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_1938_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_16_perm_0 = const()[name = string("q_with_bias_v_16_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_151_transpose_x_0 = const()[name = string("x_151_transpose_x_0"), val = bool(false)];
            bool x_151_transpose_y_0 = const()[name = string("x_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_1940_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186163968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186285888))))[name = string("op_1940_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_16_cast_fp16 = transpose(perm = q_with_bias_v_16_perm_0, x = var_1938_cast_fp16)[name = string("transpose_296")];
            tensor<fp16, [1, 8, 4, 119]> x_151_cast_fp16 = matmul(transpose_x = x_151_transpose_x_0, transpose_y = x_151_transpose_y_0, x = q_with_bias_v_16_cast_fp16, y = op_1940_to_fp16_palettized)[name = string("x_151_cast_fp16")];
            tensor<int32, [8]> x0_18_pad_0 = const()[name = string("x0_18_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_18_mode_0 = const()[name = string("x0_18_mode_0"), val = string("constant")];
            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_18_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x0_18_mode_0, pad = x0_18_pad_0, x = x_151_cast_fp16)[name = string("x0_18_cast_fp16")];
            tensor<int32, [4]> var_1948 = const()[name = string("op_1948"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_16_cast_fp16 = reshape(shape = var_1948, x = x0_18_cast_fp16)[name = string("x1_16_cast_fp16")];
            tensor<int32, [4]> var_1952_begin_0 = const()[name = string("op_1952_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1952_end_0 = const()[name = string("op_1952_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_1952_end_mask_0 = const()[name = string("op_1952_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_1952_cast_fp16 = slice_by_index(begin = var_1952_begin_0, end = var_1952_end_0, end_mask = var_1952_end_mask_0, x = x1_16_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<int32, [4]> var_1953 = const()[name = string("op_1953"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_16_cast_fp16 = reshape(shape = var_1953, x = var_1952_cast_fp16)[name = string("matrix_bd_16_cast_fp16")];
            bool matrix_ac_16_transpose_x_0 = const()[name = string("matrix_ac_16_transpose_x_0"), val = bool(false)];
            bool matrix_ac_16_transpose_y_0 = const()[name = string("matrix_ac_16_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_16_cast_fp16)[name = string("transpose_294")];
            tensor<fp16, [1, 8, 4, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_1936_cast_fp16)[name = string("transpose_295")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_16_cast_fp16 = matmul(transpose_x = matrix_ac_16_transpose_x_0, transpose_y = matrix_ac_16_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_16_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_16_begin_0 = const()[name = string("matrix_bd0_16_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_16_end_0 = const()[name = string("matrix_bd0_16_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_16_end_mask_0 = const()[name = string("matrix_bd0_16_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_16_cast_fp16 = slice_by_index(begin = matrix_bd0_16_begin_0, end = matrix_bd0_16_end_0, end_mask = matrix_bd0_16_end_mask_0, x = matrix_bd_16_cast_fp16)[name = string("matrix_bd0_16_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1962_cast_fp16 = add(x = matrix_ac_16_cast_fp16, y = matrix_bd0_16_cast_fp16)[name = string("op_1962_cast_fp16")];
            fp16 _inversed_scores_16_y_0_to_fp16 = const()[name = string("_inversed_scores_16_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_16_cast_fp16 = mul(x = var_1962_cast_fp16, y = _inversed_scores_16_y_0_to_fp16)[name = string("_inversed_scores_16_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_16_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_16_cast_fp16, cond = mask0_4)[name = string("scores0_16_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_1968_cast_fp16 = softmax(axis = var_56, x = scores0_16_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_95_cast_fp16 = select(a = var_30_to_fp16, b = var_1968_cast_fp16, cond = mask0_4)[name = string("input0_95_cast_fp16")];
            bool x2_16_transpose_x_0 = const()[name = string("x2_16_transpose_x_0"), val = bool(false)];
            bool x2_16_transpose_y_0 = const()[name = string("x2_16_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_18_cast_fp16 = transpose(perm = value_18_perm_0, x = v_16_cast_fp16)[name = string("transpose_293")];
            tensor<fp16, [1, 8, 4, 128]> x2_16_cast_fp16 = matmul(transpose_x = x2_16_transpose_x_0, transpose_y = x2_16_transpose_y_0, x = input0_95_cast_fp16, y = value_18_cast_fp16)[name = string("x2_16_cast_fp16")];
            tensor<int32, [4]> var_1972_perm_0 = const()[name = string("op_1972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1973 = const()[name = string("op_1973"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_1972_cast_fp16 = transpose(perm = var_1972_perm_0, x = x2_16_cast_fp16)[name = string("transpose_292")];
            tensor<fp16, [1, 4, 1024]> input1_48_cast_fp16 = reshape(shape = var_1973, x = var_1972_cast_fp16)[name = string("input1_48_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186286464))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187335104))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input1_48_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_97_cast_fp16 = add(x = input_97_cast_fp16, y = linear_70_cast_fp16)[name = string("input0_97_cast_fp16")];
            tensor<int32, [1]> x_155_axes_0 = const()[name = string("x_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187335680)))];
            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187337792)))];
            tensor<fp16, [1, 4, 1024]> x_155_cast_fp16 = layer_norm(axes = x_155_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input0_97_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_99_pad_type_0 = const()[name = string("input0_99_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_99_strides_0 = const()[name = string("input0_99_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_99_pad_0 = const()[name = string("input0_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_99_dilations_0 = const()[name = string("input0_99_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_99_groups_0 = const()[name = string("input0_99_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187339904))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189437120))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_155_cast_fp16)[name = string("transpose_291")];
            tensor<fp16, [1, 2048, 4]> input0_99_cast_fp16 = conv(dilations = input0_99_dilations_0, groups = input0_99_groups_0, pad = input0_99_pad_0, pad_type = input0_99_pad_type_0, strides = input0_99_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_101_cast_fp16)[name = string("input0_99_cast_fp16")];
            int32 x_157_split_num_splits_0 = const()[name = string("x_157_split_num_splits_0"), val = int32(2)];
            int32 x_157_split_axis_0 = const()[name = string("x_157_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_157_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_157_split_cast_fp16_1 = split(axis = x_157_split_axis_0, num_splits = x_157_split_num_splits_0, x = input0_99_cast_fp16)[name = string("x_157_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_157_split_1_sigmoid_cast_fp16 = sigmoid(x = x_157_split_cast_fp16_1)[name = string("x_157_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_157_cast_fp16 = mul(x = x_157_split_cast_fp16_0, y = x_157_split_1_sigmoid_cast_fp16)[name = string("x_157_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_101_cast_fp16 = select(a = var_30_to_fp16, b = x_157_cast_fp16, cond = var_570)[name = string("input0_101_cast_fp16")];
            bool new_x0_16_interleave_0 = const()[name = string("new_x0_16_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_16_cast_fp16 = concat(axis = var_56, interleave = new_x0_16_interleave_0, values = (cache14_1_cast_fp16, input0_101_cast_fp16))[name = string("new_x0_16_cast_fp16")];
            tensor<int32, [3]> var_2011_begin_0 = const()[name = string("op_2011_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2011_end_0 = const()[name = string("op_2011_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_2011_end_mask_0 = const()[name = string("op_2011_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_2011_cast_fp16 = slice_by_index(begin = var_2011_begin_0, end = var_2011_end_0, end_mask = var_2011_end_mask_0, x = new_x0_16_cast_fp16)[name = string("op_2011_cast_fp16")];
            string x_159_pad_type_0 = const()[name = string("x_159_pad_type_0"), val = string("valid")];
            int32 x_159_groups_0 = const()[name = string("x_159_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_159_strides_0 = const()[name = string("x_159_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_159_pad_0 = const()[name = string("x_159_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_159_dilations_0 = const()[name = string("x_159_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189437696))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189446976))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_159_cast_fp16 = conv(dilations = x_159_dilations_0, groups = x_159_groups_0, pad = x_159_pad_0, pad_type = x_159_pad_type_0, strides = x_159_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_16_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [3]> input1_50_perm_0 = const()[name = string("input1_50_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_161_axes_0 = const()[name = string("x_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189447552)))];
            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189449664)))];
            tensor<fp16, [1, 4, 1024]> input1_50_cast_fp16 = transpose(perm = input1_50_perm_0, x = x_159_cast_fp16)[name = string("transpose_290")];
            tensor<fp16, [1, 4, 1024]> x_161_cast_fp16 = layer_norm(axes = x_161_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input1_50_cast_fp16)[name = string("x_161_cast_fp16")];
            tensor<int32, [3]> input2_32_perm_0 = const()[name = string("input2_32_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_32_cast_fp16 = transpose(perm = input2_32_perm_0, x = x_161_cast_fp16)[name = string("transpose_289")];
            tensor<fp16, [1, 1024, 4]> var_2026_cast_fp16 = silu(x = input2_32_cast_fp16)[name = string("op_2026_cast_fp16")];
            string x_163_pad_type_0 = const()[name = string("x_163_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_163_strides_0 = const()[name = string("x_163_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_163_pad_0 = const()[name = string("x_163_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_163_dilations_0 = const()[name = string("x_163_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_163_groups_0 = const()[name = string("x_163_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189451776))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190500416))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_163_cast_fp16 = conv(dilations = x_163_dilations_0, groups = x_163_groups_0, pad = x_163_pad_0, pad_type = x_163_pad_type_0, strides = x_163_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_2026_cast_fp16)[name = string("x_163_cast_fp16")];
            tensor<int32, [3]> input3_18_perm_0 = const()[name = string("input3_18_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_18_cast_fp16 = transpose(perm = input3_18_perm_0, x = x_163_cast_fp16)[name = string("transpose_288")];
            tensor<fp16, [1, 4, 1024]> input1_52_cast_fp16 = add(x = input0_97_cast_fp16, y = input3_18_cast_fp16)[name = string("input1_52_cast_fp16")];
            tensor<int32, [1]> input0_103_axes_0 = const()[name = string("input0_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190500992)))];
            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190503104)))];
            tensor<fp16, [1, 4, 1024]> input0_103_cast_fp16 = layer_norm(axes = input0_103_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input1_52_cast_fp16)[name = string("input0_103_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190505216))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194699584))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_103_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2047_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("op_2047_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194700160))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198894528))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = var_2047_cast_fp16)[name = string("linear_72_cast_fp16")];
            fp16 var_2052_to_fp16 = const()[name = string("op_2052_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2053_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2052_to_fp16)[name = string("op_2053_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_34_cast_fp16 = add(x = input1_52_cast_fp16, y = var_2053_cast_fp16)[name = string("input2_34_cast_fp16")];
            tensor<int32, [1]> input0_105_axes_0 = const()[name = string("input0_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198895104)))];
            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198897216)))];
            tensor<fp16, [1, 4, 1024]> input0_105_cast_fp16 = layer_norm(axes = input0_105_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input2_34_cast_fp16)[name = string("input0_105_cast_fp16")];
            tensor<int32, [4]> cache15_1_begin_0 = const()[name = string("cache15_1_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> cache15_1_end_0 = const()[name = string("cache15_1_end_0"), val = tensor<int32, [4]>([9, 1, 56, 1024])];
            tensor<bool, [4]> cache15_1_end_mask_0 = const()[name = string("cache15_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache15_1_squeeze_mask_0 = const()[name = string("cache15_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache15_1_cast_fp16 = slice_by_index(begin = cache15_1_begin_0, end = cache15_1_end_0, end_mask = cache15_1_end_mask_0, squeeze_mask = cache15_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache15_1_cast_fp16")];
            tensor<int32, [4]> cache16_1_begin_0 = const()[name = string("cache16_1_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> cache16_1_end_0 = const()[name = string("cache16_1_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
            tensor<bool, [4]> cache16_1_end_mask_0 = const()[name = string("cache16_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache16_1_squeeze_mask_0 = const()[name = string("cache16_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache16_1_cast_fp16 = slice_by_index(begin = cache16_1_begin_0, end = cache16_1_end_0, end_mask = cache16_1_end_mask_0, squeeze_mask = cache16_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache16_1_cast_fp16")];
            tensor<int32, [1]> input_105_axes_0 = const()[name = string("input_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198899328)))];
            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198901440)))];
            tensor<fp16, [1, 4, 1024]> input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input0_105_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198903552))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203097920))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2082_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203098496))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207292864))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = var_2082_cast_fp16)[name = string("linear_74_cast_fp16")];
            fp16 var_2087_to_fp16 = const()[name = string("op_2087_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2088_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2087_to_fp16)[name = string("op_2088_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_109_cast_fp16 = add(x = input0_105_cast_fp16, y = var_2088_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<int32, [1]> key_18_axes_0 = const()[name = string("key_18_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207293440)))];
            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207295552)))];
            tensor<fp16, [1, 4, 1024]> key_18_cast_fp16 = layer_norm(axes = key_18_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_109_cast_fp16)[name = string("key_18_cast_fp16")];
            bool input_111_interleave_0 = const()[name = string("input_111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_111_cast_fp16 = concat(axis = var_64, interleave = input_111_interleave_0, values = (cache15_1_cast_fp16, key_18_cast_fp16))[name = string("input_111_cast_fp16")];
            tensor<int32, [3]> var_2110_begin_0 = const()[name = string("op_2110_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2110_end_0 = const()[name = string("op_2110_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_2110_end_mask_0 = const()[name = string("op_2110_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_2110_cast_fp16 = slice_by_index(begin = var_2110_begin_0, end = var_2110_end_0, end_mask = var_2110_end_mask_0, x = cache15_1_cast_fp16)[name = string("op_2110_cast_fp16")];
            bool var_2116_interleave_0 = const()[name = string("op_2116_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_2116_cast_fp16 = concat(axis = var_64, interleave = var_2116_interleave_0, values = (var_2110_cast_fp16, key_18_cast_fp16))[name = string("op_2116_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207297664))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208346304))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_18_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_2120 = const()[name = string("op_2120"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_18_cast_fp16 = reshape(shape = var_2120, x = linear_75_cast_fp16)[name = string("q_18_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208346880))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209395520))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<int32, [4]> var_2124 = const()[name = string("op_2124"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_18_cast_fp16 = reshape(shape = var_2124, x = linear_76_cast_fp16)[name = string("k_18_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209396096))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210444736))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_2128 = const()[name = string("op_2128"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_18_cast_fp16 = reshape(shape = var_2128, x = linear_77_cast_fp16)[name = string("v_18_cast_fp16")];
            tensor<int32, [4]> value_20_perm_0 = const()[name = string("value_20_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210445312)))];
            tensor<fp16, [1, 4, 8, 128]> var_2140_cast_fp16 = add(x = q_18_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210447424)))];
            tensor<fp16, [1, 4, 8, 128]> var_2142_cast_fp16 = add(x = q_18_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2142_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_18_perm_0 = const()[name = string("q_with_bias_v_18_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_171_transpose_x_0 = const()[name = string("x_171_transpose_x_0"), val = bool(false)];
            bool x_171_transpose_y_0 = const()[name = string("x_171_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_2144_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210449536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210571456))))[name = string("op_2144_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_18_cast_fp16 = transpose(perm = q_with_bias_v_18_perm_0, x = var_2142_cast_fp16)[name = string("transpose_287")];
            tensor<fp16, [1, 8, 4, 119]> x_171_cast_fp16 = matmul(transpose_x = x_171_transpose_x_0, transpose_y = x_171_transpose_y_0, x = q_with_bias_v_18_cast_fp16, y = op_2144_to_fp16_palettized)[name = string("x_171_cast_fp16")];
            tensor<int32, [8]> x0_20_pad_0 = const()[name = string("x0_20_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_20_mode_0 = const()[name = string("x0_20_mode_0"), val = string("constant")];
            fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_20_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x0_20_mode_0, pad = x0_20_pad_0, x = x_171_cast_fp16)[name = string("x0_20_cast_fp16")];
            tensor<int32, [4]> var_2152 = const()[name = string("op_2152"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_18_cast_fp16 = reshape(shape = var_2152, x = x0_20_cast_fp16)[name = string("x1_18_cast_fp16")];
            tensor<int32, [4]> var_2156_begin_0 = const()[name = string("op_2156_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2156_end_0 = const()[name = string("op_2156_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_2156_end_mask_0 = const()[name = string("op_2156_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_2156_cast_fp16 = slice_by_index(begin = var_2156_begin_0, end = var_2156_end_0, end_mask = var_2156_end_mask_0, x = x1_18_cast_fp16)[name = string("op_2156_cast_fp16")];
            tensor<int32, [4]> var_2157 = const()[name = string("op_2157"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_18_cast_fp16 = reshape(shape = var_2157, x = var_2156_cast_fp16)[name = string("matrix_bd_18_cast_fp16")];
            bool matrix_ac_18_transpose_x_0 = const()[name = string("matrix_ac_18_transpose_x_0"), val = bool(false)];
            bool matrix_ac_18_transpose_y_0 = const()[name = string("matrix_ac_18_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_18_cast_fp16)[name = string("transpose_285")];
            tensor<fp16, [1, 8, 4, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2140_cast_fp16)[name = string("transpose_286")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_18_cast_fp16 = matmul(transpose_x = matrix_ac_18_transpose_x_0, transpose_y = matrix_ac_18_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_18_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_18_begin_0 = const()[name = string("matrix_bd0_18_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_18_end_0 = const()[name = string("matrix_bd0_18_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_18_end_mask_0 = const()[name = string("matrix_bd0_18_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_18_cast_fp16 = slice_by_index(begin = matrix_bd0_18_begin_0, end = matrix_bd0_18_end_0, end_mask = matrix_bd0_18_end_mask_0, x = matrix_bd_18_cast_fp16)[name = string("matrix_bd0_18_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2166_cast_fp16 = add(x = matrix_ac_18_cast_fp16, y = matrix_bd0_18_cast_fp16)[name = string("op_2166_cast_fp16")];
            fp16 _inversed_scores_18_y_0_to_fp16 = const()[name = string("_inversed_scores_18_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_18_cast_fp16 = mul(x = var_2166_cast_fp16, y = _inversed_scores_18_y_0_to_fp16)[name = string("_inversed_scores_18_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_18_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_18_cast_fp16, cond = mask0_4)[name = string("scores0_18_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2172_cast_fp16 = softmax(axis = var_56, x = scores0_18_cast_fp16)[name = string("op_2172_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_107_cast_fp16 = select(a = var_30_to_fp16, b = var_2172_cast_fp16, cond = mask0_4)[name = string("input0_107_cast_fp16")];
            bool x2_18_transpose_x_0 = const()[name = string("x2_18_transpose_x_0"), val = bool(false)];
            bool x2_18_transpose_y_0 = const()[name = string("x2_18_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_20_cast_fp16 = transpose(perm = value_20_perm_0, x = v_18_cast_fp16)[name = string("transpose_284")];
            tensor<fp16, [1, 8, 4, 128]> x2_18_cast_fp16 = matmul(transpose_x = x2_18_transpose_x_0, transpose_y = x2_18_transpose_y_0, x = input0_107_cast_fp16, y = value_20_cast_fp16)[name = string("x2_18_cast_fp16")];
            tensor<int32, [4]> var_2176_perm_0 = const()[name = string("op_2176_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2177 = const()[name = string("op_2177"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_2176_cast_fp16 = transpose(perm = var_2176_perm_0, x = x2_18_cast_fp16)[name = string("transpose_283")];
            tensor<fp16, [1, 4, 1024]> input1_54_cast_fp16 = reshape(shape = var_2177, x = var_2176_cast_fp16)[name = string("input1_54_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210572032))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211620672))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input1_54_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_109_cast_fp16 = add(x = input_109_cast_fp16, y = linear_79_cast_fp16)[name = string("input0_109_cast_fp16")];
            tensor<int32, [1]> x_175_axes_0 = const()[name = string("x_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211621248)))];
            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211623360)))];
            tensor<fp16, [1, 4, 1024]> x_175_cast_fp16 = layer_norm(axes = x_175_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input0_109_cast_fp16)[name = string("x_175_cast_fp16")];
            tensor<int32, [3]> input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_111_pad_type_0 = const()[name = string("input0_111_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_111_strides_0 = const()[name = string("input0_111_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_111_pad_0 = const()[name = string("input0_111_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_111_dilations_0 = const()[name = string("input0_111_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_111_groups_0 = const()[name = string("input0_111_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211625472))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213722688))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_175_cast_fp16)[name = string("transpose_282")];
            tensor<fp16, [1, 2048, 4]> input0_111_cast_fp16 = conv(dilations = input0_111_dilations_0, groups = input0_111_groups_0, pad = input0_111_pad_0, pad_type = input0_111_pad_type_0, strides = input0_111_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = string("input0_111_cast_fp16")];
            int32 x_177_split_num_splits_0 = const()[name = string("x_177_split_num_splits_0"), val = int32(2)];
            int32 x_177_split_axis_0 = const()[name = string("x_177_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_177_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_177_split_cast_fp16_1 = split(axis = x_177_split_axis_0, num_splits = x_177_split_num_splits_0, x = input0_111_cast_fp16)[name = string("x_177_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_177_split_1_sigmoid_cast_fp16 = sigmoid(x = x_177_split_cast_fp16_1)[name = string("x_177_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_177_cast_fp16 = mul(x = x_177_split_cast_fp16_0, y = x_177_split_1_sigmoid_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_113_cast_fp16 = select(a = var_30_to_fp16, b = x_177_cast_fp16, cond = var_570)[name = string("input0_113_cast_fp16")];
            bool new_x0_18_interleave_0 = const()[name = string("new_x0_18_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_18_cast_fp16 = concat(axis = var_56, interleave = new_x0_18_interleave_0, values = (cache16_1_cast_fp16, input0_113_cast_fp16))[name = string("new_x0_18_cast_fp16")];
            tensor<int32, [3]> var_2215_begin_0 = const()[name = string("op_2215_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2215_end_0 = const()[name = string("op_2215_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_2215_end_mask_0 = const()[name = string("op_2215_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_2215_cast_fp16 = slice_by_index(begin = var_2215_begin_0, end = var_2215_end_0, end_mask = var_2215_end_mask_0, x = new_x0_18_cast_fp16)[name = string("op_2215_cast_fp16")];
            string x_179_pad_type_0 = const()[name = string("x_179_pad_type_0"), val = string("valid")];
            int32 x_179_groups_0 = const()[name = string("x_179_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_179_strides_0 = const()[name = string("x_179_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_179_pad_0 = const()[name = string("x_179_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_179_dilations_0 = const()[name = string("x_179_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213723264))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213732544))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_179_cast_fp16 = conv(dilations = x_179_dilations_0, groups = x_179_groups_0, pad = x_179_pad_0, pad_type = x_179_pad_type_0, strides = x_179_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_18_cast_fp16)[name = string("x_179_cast_fp16")];
            tensor<int32, [3]> input1_56_perm_0 = const()[name = string("input1_56_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_181_axes_0 = const()[name = string("x_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213733120)))];
            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213735232)))];
            tensor<fp16, [1, 4, 1024]> input1_56_cast_fp16 = transpose(perm = input1_56_perm_0, x = x_179_cast_fp16)[name = string("transpose_281")];
            tensor<fp16, [1, 4, 1024]> x_181_cast_fp16 = layer_norm(axes = x_181_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input1_56_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<int32, [3]> input2_36_perm_0 = const()[name = string("input2_36_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_36_cast_fp16 = transpose(perm = input2_36_perm_0, x = x_181_cast_fp16)[name = string("transpose_280")];
            tensor<fp16, [1, 1024, 4]> var_2230_cast_fp16 = silu(x = input2_36_cast_fp16)[name = string("op_2230_cast_fp16")];
            string x_183_pad_type_0 = const()[name = string("x_183_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_183_strides_0 = const()[name = string("x_183_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_183_pad_0 = const()[name = string("x_183_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_183_dilations_0 = const()[name = string("x_183_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_183_groups_0 = const()[name = string("x_183_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213737344))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214785984))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_183_cast_fp16 = conv(dilations = x_183_dilations_0, groups = x_183_groups_0, pad = x_183_pad_0, pad_type = x_183_pad_type_0, strides = x_183_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_2230_cast_fp16)[name = string("x_183_cast_fp16")];
            tensor<int32, [3]> input3_20_perm_0 = const()[name = string("input3_20_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_20_cast_fp16 = transpose(perm = input3_20_perm_0, x = x_183_cast_fp16)[name = string("transpose_279")];
            tensor<fp16, [1, 4, 1024]> input1_58_cast_fp16 = add(x = input0_109_cast_fp16, y = input3_20_cast_fp16)[name = string("input1_58_cast_fp16")];
            tensor<int32, [1]> input0_115_axes_0 = const()[name = string("input0_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214786560)))];
            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214788672)))];
            tensor<fp16, [1, 4, 1024]> input0_115_cast_fp16 = layer_norm(axes = input0_115_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input1_58_cast_fp16)[name = string("input0_115_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214790784))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218985152))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_115_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2251_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218985728))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223180096))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = var_2251_cast_fp16)[name = string("linear_81_cast_fp16")];
            fp16 var_2256_to_fp16 = const()[name = string("op_2256_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2257_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2256_to_fp16)[name = string("op_2257_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_38_cast_fp16 = add(x = input1_58_cast_fp16, y = var_2257_cast_fp16)[name = string("input2_38_cast_fp16")];
            tensor<int32, [1]> input0_117_axes_0 = const()[name = string("input0_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223180672)))];
            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223182784)))];
            tensor<fp16, [1, 4, 1024]> input0_117_cast_fp16 = layer_norm(axes = input0_117_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input2_38_cast_fp16)[name = string("input0_117_cast_fp16")];
            tensor<int32, [4]> cache17_1_begin_0 = const()[name = string("cache17_1_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> cache17_1_end_0 = const()[name = string("cache17_1_end_0"), val = tensor<int32, [4]>([10, 1, 56, 1024])];
            tensor<bool, [4]> cache17_1_end_mask_0 = const()[name = string("cache17_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache17_1_squeeze_mask_0 = const()[name = string("cache17_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache17_1_cast_fp16 = slice_by_index(begin = cache17_1_begin_0, end = cache17_1_end_0, end_mask = cache17_1_end_mask_0, squeeze_mask = cache17_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache17_1_cast_fp16")];
            tensor<int32, [4]> cache18_1_begin_0 = const()[name = string("cache18_1_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> cache18_1_end_0 = const()[name = string("cache18_1_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
            tensor<bool, [4]> cache18_1_end_mask_0 = const()[name = string("cache18_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache18_1_squeeze_mask_0 = const()[name = string("cache18_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache18_1_cast_fp16 = slice_by_index(begin = cache18_1_begin_0, end = cache18_1_end_0, end_mask = cache18_1_end_mask_0, squeeze_mask = cache18_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache18_1_cast_fp16")];
            tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223184896)))];
            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223187008)))];
            tensor<fp16, [1, 4, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input0_117_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223189120))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227383488))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2286_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("op_2286_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227384064))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231578432))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = var_2286_cast_fp16)[name = string("linear_83_cast_fp16")];
            fp16 var_2291_to_fp16 = const()[name = string("op_2291_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2292_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2291_to_fp16)[name = string("op_2292_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_121_cast_fp16 = add(x = input0_117_cast_fp16, y = var_2292_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<int32, [1]> key_20_axes_0 = const()[name = string("key_20_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231579008)))];
            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231581120)))];
            tensor<fp16, [1, 4, 1024]> key_20_cast_fp16 = layer_norm(axes = key_20_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_121_cast_fp16)[name = string("key_20_cast_fp16")];
            bool input_123_interleave_0 = const()[name = string("input_123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_123_cast_fp16 = concat(axis = var_64, interleave = input_123_interleave_0, values = (cache17_1_cast_fp16, key_20_cast_fp16))[name = string("input_123_cast_fp16")];
            tensor<int32, [3]> var_2314_begin_0 = const()[name = string("op_2314_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2314_end_0 = const()[name = string("op_2314_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_2314_end_mask_0 = const()[name = string("op_2314_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_2314_cast_fp16 = slice_by_index(begin = var_2314_begin_0, end = var_2314_end_0, end_mask = var_2314_end_mask_0, x = cache17_1_cast_fp16)[name = string("op_2314_cast_fp16")];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_2320_cast_fp16 = concat(axis = var_64, interleave = var_2320_interleave_0, values = (var_2314_cast_fp16, key_20_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231583232))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232631872))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_20_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_2324 = const()[name = string("op_2324"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_20_cast_fp16 = reshape(shape = var_2324, x = linear_84_cast_fp16)[name = string("q_20_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232632448))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233681088))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_2328 = const()[name = string("op_2328"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_20_cast_fp16 = reshape(shape = var_2328, x = linear_85_cast_fp16)[name = string("k_20_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233681664))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234730304))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_2332 = const()[name = string("op_2332"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_20_cast_fp16 = reshape(shape = var_2332, x = linear_86_cast_fp16)[name = string("v_20_cast_fp16")];
            tensor<int32, [4]> value_22_perm_0 = const()[name = string("value_22_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234730880)))];
            tensor<fp16, [1, 4, 8, 128]> var_2344_cast_fp16 = add(x = q_20_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2344_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234732992)))];
            tensor<fp16, [1, 4, 8, 128]> var_2346_cast_fp16 = add(x = q_20_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2346_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_20_perm_0 = const()[name = string("q_with_bias_v_20_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_191_transpose_x_0 = const()[name = string("x_191_transpose_x_0"), val = bool(false)];
            bool x_191_transpose_y_0 = const()[name = string("x_191_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_2348_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234735104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234857024))))[name = string("op_2348_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_20_cast_fp16 = transpose(perm = q_with_bias_v_20_perm_0, x = var_2346_cast_fp16)[name = string("transpose_278")];
            tensor<fp16, [1, 8, 4, 119]> x_191_cast_fp16 = matmul(transpose_x = x_191_transpose_x_0, transpose_y = x_191_transpose_y_0, x = q_with_bias_v_20_cast_fp16, y = op_2348_to_fp16_palettized)[name = string("x_191_cast_fp16")];
            tensor<int32, [8]> x0_22_pad_0 = const()[name = string("x0_22_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_22_mode_0 = const()[name = string("x0_22_mode_0"), val = string("constant")];
            fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_22_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x0_22_mode_0, pad = x0_22_pad_0, x = x_191_cast_fp16)[name = string("x0_22_cast_fp16")];
            tensor<int32, [4]> var_2356 = const()[name = string("op_2356"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_20_cast_fp16 = reshape(shape = var_2356, x = x0_22_cast_fp16)[name = string("x1_20_cast_fp16")];
            tensor<int32, [4]> var_2360_begin_0 = const()[name = string("op_2360_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2360_end_0 = const()[name = string("op_2360_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_2360_end_mask_0 = const()[name = string("op_2360_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_2360_cast_fp16 = slice_by_index(begin = var_2360_begin_0, end = var_2360_end_0, end_mask = var_2360_end_mask_0, x = x1_20_cast_fp16)[name = string("op_2360_cast_fp16")];
            tensor<int32, [4]> var_2361 = const()[name = string("op_2361"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_20_cast_fp16 = reshape(shape = var_2361, x = var_2360_cast_fp16)[name = string("matrix_bd_20_cast_fp16")];
            bool matrix_ac_20_transpose_x_0 = const()[name = string("matrix_ac_20_transpose_x_0"), val = bool(false)];
            bool matrix_ac_20_transpose_y_0 = const()[name = string("matrix_ac_20_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_20_cast_fp16)[name = string("transpose_276")];
            tensor<fp16, [1, 8, 4, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2344_cast_fp16)[name = string("transpose_277")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_20_cast_fp16 = matmul(transpose_x = matrix_ac_20_transpose_x_0, transpose_y = matrix_ac_20_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_20_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_20_begin_0 = const()[name = string("matrix_bd0_20_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_20_end_0 = const()[name = string("matrix_bd0_20_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_20_end_mask_0 = const()[name = string("matrix_bd0_20_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_20_cast_fp16 = slice_by_index(begin = matrix_bd0_20_begin_0, end = matrix_bd0_20_end_0, end_mask = matrix_bd0_20_end_mask_0, x = matrix_bd_20_cast_fp16)[name = string("matrix_bd0_20_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2370_cast_fp16 = add(x = matrix_ac_20_cast_fp16, y = matrix_bd0_20_cast_fp16)[name = string("op_2370_cast_fp16")];
            fp16 _inversed_scores_20_y_0_to_fp16 = const()[name = string("_inversed_scores_20_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_20_cast_fp16 = mul(x = var_2370_cast_fp16, y = _inversed_scores_20_y_0_to_fp16)[name = string("_inversed_scores_20_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_20_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_20_cast_fp16, cond = mask0_4)[name = string("scores0_20_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2376_cast_fp16 = softmax(axis = var_56, x = scores0_20_cast_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_119_cast_fp16 = select(a = var_30_to_fp16, b = var_2376_cast_fp16, cond = mask0_4)[name = string("input0_119_cast_fp16")];
            bool x2_20_transpose_x_0 = const()[name = string("x2_20_transpose_x_0"), val = bool(false)];
            bool x2_20_transpose_y_0 = const()[name = string("x2_20_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_22_cast_fp16 = transpose(perm = value_22_perm_0, x = v_20_cast_fp16)[name = string("transpose_275")];
            tensor<fp16, [1, 8, 4, 128]> x2_20_cast_fp16 = matmul(transpose_x = x2_20_transpose_x_0, transpose_y = x2_20_transpose_y_0, x = input0_119_cast_fp16, y = value_22_cast_fp16)[name = string("x2_20_cast_fp16")];
            tensor<int32, [4]> var_2380_perm_0 = const()[name = string("op_2380_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2381 = const()[name = string("op_2381"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_2380_cast_fp16 = transpose(perm = var_2380_perm_0, x = x2_20_cast_fp16)[name = string("transpose_274")];
            tensor<fp16, [1, 4, 1024]> input1_60_cast_fp16 = reshape(shape = var_2381, x = var_2380_cast_fp16)[name = string("input1_60_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234857600))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235906240))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input1_60_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_121_cast_fp16 = add(x = input_121_cast_fp16, y = linear_88_cast_fp16)[name = string("input0_121_cast_fp16")];
            tensor<int32, [1]> x_195_axes_0 = const()[name = string("x_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235906816)))];
            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235908928)))];
            tensor<fp16, [1, 4, 1024]> x_195_cast_fp16 = layer_norm(axes = x_195_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input0_121_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [3]> input_125_perm_0 = const()[name = string("input_125_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_123_pad_type_0 = const()[name = string("input0_123_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_123_strides_0 = const()[name = string("input0_123_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_123_pad_0 = const()[name = string("input0_123_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_123_dilations_0 = const()[name = string("input0_123_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_123_groups_0 = const()[name = string("input0_123_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235911040))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238008256))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_125_cast_fp16 = transpose(perm = input_125_perm_0, x = x_195_cast_fp16)[name = string("transpose_273")];
            tensor<fp16, [1, 2048, 4]> input0_123_cast_fp16 = conv(dilations = input0_123_dilations_0, groups = input0_123_groups_0, pad = input0_123_pad_0, pad_type = input0_123_pad_type_0, strides = input0_123_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = string("input0_123_cast_fp16")];
            int32 x_197_split_num_splits_0 = const()[name = string("x_197_split_num_splits_0"), val = int32(2)];
            int32 x_197_split_axis_0 = const()[name = string("x_197_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_197_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_197_split_cast_fp16_1 = split(axis = x_197_split_axis_0, num_splits = x_197_split_num_splits_0, x = input0_123_cast_fp16)[name = string("x_197_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_197_split_1_sigmoid_cast_fp16 = sigmoid(x = x_197_split_cast_fp16_1)[name = string("x_197_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_197_cast_fp16 = mul(x = x_197_split_cast_fp16_0, y = x_197_split_1_sigmoid_cast_fp16)[name = string("x_197_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_125_cast_fp16 = select(a = var_30_to_fp16, b = x_197_cast_fp16, cond = var_570)[name = string("input0_125_cast_fp16")];
            bool new_x0_20_interleave_0 = const()[name = string("new_x0_20_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_20_cast_fp16 = concat(axis = var_56, interleave = new_x0_20_interleave_0, values = (cache18_1_cast_fp16, input0_125_cast_fp16))[name = string("new_x0_20_cast_fp16")];
            tensor<int32, [3]> var_2419_begin_0 = const()[name = string("op_2419_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2419_end_0 = const()[name = string("op_2419_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_2419_end_mask_0 = const()[name = string("op_2419_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_2419_cast_fp16 = slice_by_index(begin = var_2419_begin_0, end = var_2419_end_0, end_mask = var_2419_end_mask_0, x = new_x0_20_cast_fp16)[name = string("op_2419_cast_fp16")];
            string x_199_pad_type_0 = const()[name = string("x_199_pad_type_0"), val = string("valid")];
            int32 x_199_groups_0 = const()[name = string("x_199_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_199_strides_0 = const()[name = string("x_199_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_199_pad_0 = const()[name = string("x_199_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_199_dilations_0 = const()[name = string("x_199_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238008832))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238018112))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_199_cast_fp16 = conv(dilations = x_199_dilations_0, groups = x_199_groups_0, pad = x_199_pad_0, pad_type = x_199_pad_type_0, strides = x_199_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_20_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<int32, [3]> input1_62_perm_0 = const()[name = string("input1_62_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_201_axes_0 = const()[name = string("x_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238018688)))];
            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238020800)))];
            tensor<fp16, [1, 4, 1024]> input1_62_cast_fp16 = transpose(perm = input1_62_perm_0, x = x_199_cast_fp16)[name = string("transpose_272")];
            tensor<fp16, [1, 4, 1024]> x_201_cast_fp16 = layer_norm(axes = x_201_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input1_62_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<int32, [3]> input2_40_perm_0 = const()[name = string("input2_40_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_40_cast_fp16 = transpose(perm = input2_40_perm_0, x = x_201_cast_fp16)[name = string("transpose_271")];
            tensor<fp16, [1, 1024, 4]> var_2434_cast_fp16 = silu(x = input2_40_cast_fp16)[name = string("op_2434_cast_fp16")];
            string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238022912))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239071552))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_2434_cast_fp16)[name = string("x_203_cast_fp16")];
            tensor<int32, [3]> input3_22_perm_0 = const()[name = string("input3_22_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_22_cast_fp16 = transpose(perm = input3_22_perm_0, x = x_203_cast_fp16)[name = string("transpose_270")];
            tensor<fp16, [1, 4, 1024]> input1_64_cast_fp16 = add(x = input0_121_cast_fp16, y = input3_22_cast_fp16)[name = string("input1_64_cast_fp16")];
            tensor<int32, [1]> input0_127_axes_0 = const()[name = string("input0_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239072128)))];
            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239074240)))];
            tensor<fp16, [1, 4, 1024]> input0_127_cast_fp16 = layer_norm(axes = input0_127_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input1_64_cast_fp16)[name = string("input0_127_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239076352))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243270720))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_127_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2455_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("op_2455_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243271296))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247465664))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = var_2455_cast_fp16)[name = string("linear_90_cast_fp16")];
            fp16 var_2460_to_fp16 = const()[name = string("op_2460_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2461_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2460_to_fp16)[name = string("op_2461_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_42_cast_fp16 = add(x = input1_64_cast_fp16, y = var_2461_cast_fp16)[name = string("input2_42_cast_fp16")];
            tensor<int32, [1]> input0_129_axes_0 = const()[name = string("input0_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247466240)))];
            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247468352)))];
            tensor<fp16, [1, 4, 1024]> input0_129_cast_fp16 = layer_norm(axes = input0_129_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input2_42_cast_fp16)[name = string("input0_129_cast_fp16")];
            tensor<int32, [4]> cache19_1_begin_0 = const()[name = string("cache19_1_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> cache19_1_end_0 = const()[name = string("cache19_1_end_0"), val = tensor<int32, [4]>([11, 1, 56, 1024])];
            tensor<bool, [4]> cache19_1_end_mask_0 = const()[name = string("cache19_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache19_1_squeeze_mask_0 = const()[name = string("cache19_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache19_1_cast_fp16 = slice_by_index(begin = cache19_1_begin_0, end = cache19_1_end_0, end_mask = cache19_1_end_mask_0, squeeze_mask = cache19_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache19_1_cast_fp16")];
            tensor<int32, [4]> cache20_1_begin_0 = const()[name = string("cache20_1_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> cache20_1_end_0 = const()[name = string("cache20_1_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
            tensor<bool, [4]> cache20_1_end_mask_0 = const()[name = string("cache20_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache20_1_squeeze_mask_0 = const()[name = string("cache20_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache20_1_cast_fp16 = slice_by_index(begin = cache20_1_begin_0, end = cache20_1_end_0, end_mask = cache20_1_end_mask_0, squeeze_mask = cache20_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache20_1_cast_fp16")];
            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247470464)))];
            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247472576)))];
            tensor<fp16, [1, 4, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input0_129_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247474688))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251669056))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2490_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251669632))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255864000))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = var_2490_cast_fp16)[name = string("linear_92_cast_fp16")];
            fp16 var_2495_to_fp16 = const()[name = string("op_2495_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2496_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2495_to_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_133_cast_fp16 = add(x = input0_129_cast_fp16, y = var_2496_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<int32, [1]> key_22_axes_0 = const()[name = string("key_22_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255864576)))];
            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255866688)))];
            tensor<fp16, [1, 4, 1024]> key_22_cast_fp16 = layer_norm(axes = key_22_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_133_cast_fp16)[name = string("key_22_cast_fp16")];
            bool input_135_interleave_0 = const()[name = string("input_135_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_135_cast_fp16 = concat(axis = var_64, interleave = input_135_interleave_0, values = (cache19_1_cast_fp16, key_22_cast_fp16))[name = string("input_135_cast_fp16")];
            tensor<int32, [3]> var_2518_begin_0 = const()[name = string("op_2518_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2518_end_0 = const()[name = string("op_2518_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_2518_end_mask_0 = const()[name = string("op_2518_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_2518_cast_fp16 = slice_by_index(begin = var_2518_begin_0, end = var_2518_end_0, end_mask = var_2518_end_mask_0, x = cache19_1_cast_fp16)[name = string("op_2518_cast_fp16")];
            bool var_2524_interleave_0 = const()[name = string("op_2524_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_2524_cast_fp16 = concat(axis = var_64, interleave = var_2524_interleave_0, values = (var_2518_cast_fp16, key_22_cast_fp16))[name = string("op_2524_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255868800))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256917440))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_22_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2528 = const()[name = string("op_2528"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_22_cast_fp16 = reshape(shape = var_2528, x = linear_93_cast_fp16)[name = string("q_22_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256918016))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257966656))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [4]> var_2532 = const()[name = string("op_2532"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_22_cast_fp16 = reshape(shape = var_2532, x = linear_94_cast_fp16)[name = string("k_22_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257967232))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259015872))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<int32, [4]> var_2536 = const()[name = string("op_2536"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_22_cast_fp16 = reshape(shape = var_2536, x = linear_95_cast_fp16)[name = string("v_22_cast_fp16")];
            tensor<int32, [4]> value_24_perm_0 = const()[name = string("value_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259016448)))];
            tensor<fp16, [1, 4, 8, 128]> var_2548_cast_fp16 = add(x = q_22_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2548_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259018560)))];
            tensor<fp16, [1, 4, 8, 128]> var_2550_cast_fp16 = add(x = q_22_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2550_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_22_perm_0 = const()[name = string("q_with_bias_v_22_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_211_transpose_x_0 = const()[name = string("x_211_transpose_x_0"), val = bool(false)];
            bool x_211_transpose_y_0 = const()[name = string("x_211_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_2552_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259020672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259142592))))[name = string("op_2552_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_22_cast_fp16 = transpose(perm = q_with_bias_v_22_perm_0, x = var_2550_cast_fp16)[name = string("transpose_269")];
            tensor<fp16, [1, 8, 4, 119]> x_211_cast_fp16 = matmul(transpose_x = x_211_transpose_x_0, transpose_y = x_211_transpose_y_0, x = q_with_bias_v_22_cast_fp16, y = op_2552_to_fp16_palettized)[name = string("x_211_cast_fp16")];
            tensor<int32, [8]> x0_24_pad_0 = const()[name = string("x0_24_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_24_mode_0 = const()[name = string("x0_24_mode_0"), val = string("constant")];
            fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_24_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x0_24_mode_0, pad = x0_24_pad_0, x = x_211_cast_fp16)[name = string("x0_24_cast_fp16")];
            tensor<int32, [4]> var_2560 = const()[name = string("op_2560"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_22_cast_fp16 = reshape(shape = var_2560, x = x0_24_cast_fp16)[name = string("x1_22_cast_fp16")];
            tensor<int32, [4]> var_2564_begin_0 = const()[name = string("op_2564_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2564_end_0 = const()[name = string("op_2564_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_2564_end_mask_0 = const()[name = string("op_2564_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_2564_cast_fp16 = slice_by_index(begin = var_2564_begin_0, end = var_2564_end_0, end_mask = var_2564_end_mask_0, x = x1_22_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> var_2565 = const()[name = string("op_2565"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_22_cast_fp16 = reshape(shape = var_2565, x = var_2564_cast_fp16)[name = string("matrix_bd_22_cast_fp16")];
            bool matrix_ac_22_transpose_x_0 = const()[name = string("matrix_ac_22_transpose_x_0"), val = bool(false)];
            bool matrix_ac_22_transpose_y_0 = const()[name = string("matrix_ac_22_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_22_cast_fp16)[name = string("transpose_267")];
            tensor<fp16, [1, 8, 4, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2548_cast_fp16)[name = string("transpose_268")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_22_cast_fp16 = matmul(transpose_x = matrix_ac_22_transpose_x_0, transpose_y = matrix_ac_22_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_22_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_22_begin_0 = const()[name = string("matrix_bd0_22_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_22_end_0 = const()[name = string("matrix_bd0_22_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_22_end_mask_0 = const()[name = string("matrix_bd0_22_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_22_cast_fp16 = slice_by_index(begin = matrix_bd0_22_begin_0, end = matrix_bd0_22_end_0, end_mask = matrix_bd0_22_end_mask_0, x = matrix_bd_22_cast_fp16)[name = string("matrix_bd0_22_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2574_cast_fp16 = add(x = matrix_ac_22_cast_fp16, y = matrix_bd0_22_cast_fp16)[name = string("op_2574_cast_fp16")];
            fp16 _inversed_scores_22_y_0_to_fp16 = const()[name = string("_inversed_scores_22_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_22_cast_fp16 = mul(x = var_2574_cast_fp16, y = _inversed_scores_22_y_0_to_fp16)[name = string("_inversed_scores_22_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_22_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_22_cast_fp16, cond = mask0_4)[name = string("scores0_22_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2580_cast_fp16 = softmax(axis = var_56, x = scores0_22_cast_fp16)[name = string("op_2580_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_131_cast_fp16 = select(a = var_30_to_fp16, b = var_2580_cast_fp16, cond = mask0_4)[name = string("input0_131_cast_fp16")];
            bool x2_22_transpose_x_0 = const()[name = string("x2_22_transpose_x_0"), val = bool(false)];
            bool x2_22_transpose_y_0 = const()[name = string("x2_22_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_24_cast_fp16 = transpose(perm = value_24_perm_0, x = v_22_cast_fp16)[name = string("transpose_266")];
            tensor<fp16, [1, 8, 4, 128]> x2_22_cast_fp16 = matmul(transpose_x = x2_22_transpose_x_0, transpose_y = x2_22_transpose_y_0, x = input0_131_cast_fp16, y = value_24_cast_fp16)[name = string("x2_22_cast_fp16")];
            tensor<int32, [4]> var_2584_perm_0 = const()[name = string("op_2584_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2585 = const()[name = string("op_2585"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_2584_cast_fp16 = transpose(perm = var_2584_perm_0, x = x2_22_cast_fp16)[name = string("transpose_265")];
            tensor<fp16, [1, 4, 1024]> input1_66_cast_fp16 = reshape(shape = var_2585, x = var_2584_cast_fp16)[name = string("input1_66_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259143168))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260191808))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input1_66_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_133_cast_fp16 = add(x = input_133_cast_fp16, y = linear_97_cast_fp16)[name = string("input0_133_cast_fp16")];
            tensor<int32, [1]> x_215_axes_0 = const()[name = string("x_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260192384)))];
            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260194496)))];
            tensor<fp16, [1, 4, 1024]> x_215_cast_fp16 = layer_norm(axes = x_215_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input0_133_cast_fp16)[name = string("x_215_cast_fp16")];
            tensor<int32, [3]> input_137_perm_0 = const()[name = string("input_137_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_135_pad_type_0 = const()[name = string("input0_135_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_135_strides_0 = const()[name = string("input0_135_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_135_pad_0 = const()[name = string("input0_135_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_135_dilations_0 = const()[name = string("input0_135_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_135_groups_0 = const()[name = string("input0_135_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260196608))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262293824))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_137_cast_fp16 = transpose(perm = input_137_perm_0, x = x_215_cast_fp16)[name = string("transpose_264")];
            tensor<fp16, [1, 2048, 4]> input0_135_cast_fp16 = conv(dilations = input0_135_dilations_0, groups = input0_135_groups_0, pad = input0_135_pad_0, pad_type = input0_135_pad_type_0, strides = input0_135_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_137_cast_fp16)[name = string("input0_135_cast_fp16")];
            int32 x_217_split_num_splits_0 = const()[name = string("x_217_split_num_splits_0"), val = int32(2)];
            int32 x_217_split_axis_0 = const()[name = string("x_217_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_217_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_217_split_cast_fp16_1 = split(axis = x_217_split_axis_0, num_splits = x_217_split_num_splits_0, x = input0_135_cast_fp16)[name = string("x_217_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_217_split_1_sigmoid_cast_fp16 = sigmoid(x = x_217_split_cast_fp16_1)[name = string("x_217_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_217_cast_fp16 = mul(x = x_217_split_cast_fp16_0, y = x_217_split_1_sigmoid_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_137_cast_fp16 = select(a = var_30_to_fp16, b = x_217_cast_fp16, cond = var_570)[name = string("input0_137_cast_fp16")];
            bool new_x0_22_interleave_0 = const()[name = string("new_x0_22_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_22_cast_fp16 = concat(axis = var_56, interleave = new_x0_22_interleave_0, values = (cache20_1_cast_fp16, input0_137_cast_fp16))[name = string("new_x0_22_cast_fp16")];
            tensor<int32, [3]> var_2623_begin_0 = const()[name = string("op_2623_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2623_end_0 = const()[name = string("op_2623_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_2623_end_mask_0 = const()[name = string("op_2623_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_2623_cast_fp16 = slice_by_index(begin = var_2623_begin_0, end = var_2623_end_0, end_mask = var_2623_end_mask_0, x = new_x0_22_cast_fp16)[name = string("op_2623_cast_fp16")];
            string x_219_pad_type_0 = const()[name = string("x_219_pad_type_0"), val = string("valid")];
            int32 x_219_groups_0 = const()[name = string("x_219_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_219_strides_0 = const()[name = string("x_219_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_219_pad_0 = const()[name = string("x_219_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_219_dilations_0 = const()[name = string("x_219_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262294400))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262303680))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_219_cast_fp16 = conv(dilations = x_219_dilations_0, groups = x_219_groups_0, pad = x_219_pad_0, pad_type = x_219_pad_type_0, strides = x_219_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_22_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<int32, [3]> input1_68_perm_0 = const()[name = string("input1_68_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_221_axes_0 = const()[name = string("x_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262304256)))];
            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262306368)))];
            tensor<fp16, [1, 4, 1024]> input1_68_cast_fp16 = transpose(perm = input1_68_perm_0, x = x_219_cast_fp16)[name = string("transpose_263")];
            tensor<fp16, [1, 4, 1024]> x_221_cast_fp16 = layer_norm(axes = x_221_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input1_68_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<int32, [3]> input2_44_perm_0 = const()[name = string("input2_44_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_44_cast_fp16 = transpose(perm = input2_44_perm_0, x = x_221_cast_fp16)[name = string("transpose_262")];
            tensor<fp16, [1, 1024, 4]> var_2638_cast_fp16 = silu(x = input2_44_cast_fp16)[name = string("op_2638_cast_fp16")];
            string x_223_pad_type_0 = const()[name = string("x_223_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_223_strides_0 = const()[name = string("x_223_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_223_pad_0 = const()[name = string("x_223_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_223_dilations_0 = const()[name = string("x_223_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_223_groups_0 = const()[name = string("x_223_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262308480))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263357120))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_223_cast_fp16 = conv(dilations = x_223_dilations_0, groups = x_223_groups_0, pad = x_223_pad_0, pad_type = x_223_pad_type_0, strides = x_223_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_2638_cast_fp16)[name = string("x_223_cast_fp16")];
            tensor<int32, [3]> input3_24_perm_0 = const()[name = string("input3_24_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_24_cast_fp16 = transpose(perm = input3_24_perm_0, x = x_223_cast_fp16)[name = string("transpose_261")];
            tensor<fp16, [1, 4, 1024]> input1_70_cast_fp16 = add(x = input0_133_cast_fp16, y = input3_24_cast_fp16)[name = string("input1_70_cast_fp16")];
            tensor<int32, [1]> input0_139_axes_0 = const()[name = string("input0_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263357696)))];
            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263359808)))];
            tensor<fp16, [1, 4, 1024]> input0_139_cast_fp16 = layer_norm(axes = input0_139_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input1_70_cast_fp16)[name = string("input0_139_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263361920))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267556288))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_139_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2659_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("op_2659_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267556864))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271751232))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = var_2659_cast_fp16)[name = string("linear_99_cast_fp16")];
            fp16 var_2664_to_fp16 = const()[name = string("op_2664_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2665_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2664_to_fp16)[name = string("op_2665_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_46_cast_fp16 = add(x = input1_70_cast_fp16, y = var_2665_cast_fp16)[name = string("input2_46_cast_fp16")];
            tensor<int32, [1]> input0_141_axes_0 = const()[name = string("input0_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271751808)))];
            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271753920)))];
            tensor<fp16, [1, 4, 1024]> input0_141_cast_fp16 = layer_norm(axes = input0_141_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input2_46_cast_fp16)[name = string("input0_141_cast_fp16")];
            tensor<int32, [4]> cache21_1_begin_0 = const()[name = string("cache21_1_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> cache21_1_end_0 = const()[name = string("cache21_1_end_0"), val = tensor<int32, [4]>([12, 1, 56, 1024])];
            tensor<bool, [4]> cache21_1_end_mask_0 = const()[name = string("cache21_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache21_1_squeeze_mask_0 = const()[name = string("cache21_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache21_1_cast_fp16 = slice_by_index(begin = cache21_1_begin_0, end = cache21_1_end_0, end_mask = cache21_1_end_mask_0, squeeze_mask = cache21_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache21_1_cast_fp16")];
            tensor<int32, [4]> cache22_1_begin_0 = const()[name = string("cache22_1_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> cache22_1_end_0 = const()[name = string("cache22_1_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
            tensor<bool, [4]> cache22_1_end_mask_0 = const()[name = string("cache22_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache22_1_squeeze_mask_0 = const()[name = string("cache22_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache22_1_cast_fp16 = slice_by_index(begin = cache22_1_begin_0, end = cache22_1_end_0, end_mask = cache22_1_end_mask_0, squeeze_mask = cache22_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache22_1_cast_fp16")];
            tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271756032)))];
            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271758144)))];
            tensor<fp16, [1, 4, 1024]> input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input0_141_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271760256))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275954624))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2694_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275955200))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280149568))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = var_2694_cast_fp16)[name = string("linear_101_cast_fp16")];
            fp16 var_2699_to_fp16 = const()[name = string("op_2699_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2700_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2699_to_fp16)[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_145_cast_fp16 = add(x = input0_141_cast_fp16, y = var_2700_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<int32, [1]> key_24_axes_0 = const()[name = string("key_24_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280150144)))];
            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280152256)))];
            tensor<fp16, [1, 4, 1024]> key_24_cast_fp16 = layer_norm(axes = key_24_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_145_cast_fp16)[name = string("key_24_cast_fp16")];
            bool input_147_interleave_0 = const()[name = string("input_147_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_147_cast_fp16 = concat(axis = var_64, interleave = input_147_interleave_0, values = (cache21_1_cast_fp16, key_24_cast_fp16))[name = string("input_147_cast_fp16")];
            tensor<int32, [3]> var_2722_begin_0 = const()[name = string("op_2722_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2722_end_0 = const()[name = string("op_2722_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_2722_end_mask_0 = const()[name = string("op_2722_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_2722_cast_fp16 = slice_by_index(begin = var_2722_begin_0, end = var_2722_end_0, end_mask = var_2722_end_mask_0, x = cache21_1_cast_fp16)[name = string("op_2722_cast_fp16")];
            bool var_2728_interleave_0 = const()[name = string("op_2728_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_2728_cast_fp16 = concat(axis = var_64, interleave = var_2728_interleave_0, values = (var_2722_cast_fp16, key_24_cast_fp16))[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280154368))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281203008))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_24_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_2732 = const()[name = string("op_2732"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_24_cast_fp16 = reshape(shape = var_2732, x = linear_102_cast_fp16)[name = string("q_24_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281203584))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282252224))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<int32, [4]> var_2736 = const()[name = string("op_2736"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_24_cast_fp16 = reshape(shape = var_2736, x = linear_103_cast_fp16)[name = string("k_24_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282252800))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283301440))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [4]> var_2740 = const()[name = string("op_2740"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_24_cast_fp16 = reshape(shape = var_2740, x = linear_104_cast_fp16)[name = string("v_24_cast_fp16")];
            tensor<int32, [4]> value_26_perm_0 = const()[name = string("value_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283302016)))];
            tensor<fp16, [1, 4, 8, 128]> var_2752_cast_fp16 = add(x = q_24_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2752_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283304128)))];
            tensor<fp16, [1, 4, 8, 128]> var_2754_cast_fp16 = add(x = q_24_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2754_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_24_perm_0 = const()[name = string("q_with_bias_v_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_231_transpose_x_0 = const()[name = string("x_231_transpose_x_0"), val = bool(false)];
            bool x_231_transpose_y_0 = const()[name = string("x_231_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_2756_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283306240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283428160))))[name = string("op_2756_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_24_cast_fp16 = transpose(perm = q_with_bias_v_24_perm_0, x = var_2754_cast_fp16)[name = string("transpose_260")];
            tensor<fp16, [1, 8, 4, 119]> x_231_cast_fp16 = matmul(transpose_x = x_231_transpose_x_0, transpose_y = x_231_transpose_y_0, x = q_with_bias_v_24_cast_fp16, y = op_2756_to_fp16_palettized)[name = string("x_231_cast_fp16")];
            tensor<int32, [8]> x0_26_pad_0 = const()[name = string("x0_26_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_26_mode_0 = const()[name = string("x0_26_mode_0"), val = string("constant")];
            fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_26_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x0_26_mode_0, pad = x0_26_pad_0, x = x_231_cast_fp16)[name = string("x0_26_cast_fp16")];
            tensor<int32, [4]> var_2764 = const()[name = string("op_2764"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_24_cast_fp16 = reshape(shape = var_2764, x = x0_26_cast_fp16)[name = string("x1_24_cast_fp16")];
            tensor<int32, [4]> var_2768_begin_0 = const()[name = string("op_2768_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2768_end_0 = const()[name = string("op_2768_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_2768_end_mask_0 = const()[name = string("op_2768_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_2768_cast_fp16 = slice_by_index(begin = var_2768_begin_0, end = var_2768_end_0, end_mask = var_2768_end_mask_0, x = x1_24_cast_fp16)[name = string("op_2768_cast_fp16")];
            tensor<int32, [4]> var_2769 = const()[name = string("op_2769"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_24_cast_fp16 = reshape(shape = var_2769, x = var_2768_cast_fp16)[name = string("matrix_bd_24_cast_fp16")];
            bool matrix_ac_24_transpose_x_0 = const()[name = string("matrix_ac_24_transpose_x_0"), val = bool(false)];
            bool matrix_ac_24_transpose_y_0 = const()[name = string("matrix_ac_24_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_24_cast_fp16)[name = string("transpose_258")];
            tensor<fp16, [1, 8, 4, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2752_cast_fp16)[name = string("transpose_259")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_24_cast_fp16 = matmul(transpose_x = matrix_ac_24_transpose_x_0, transpose_y = matrix_ac_24_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_24_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_24_begin_0 = const()[name = string("matrix_bd0_24_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_24_end_0 = const()[name = string("matrix_bd0_24_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_24_end_mask_0 = const()[name = string("matrix_bd0_24_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_24_cast_fp16 = slice_by_index(begin = matrix_bd0_24_begin_0, end = matrix_bd0_24_end_0, end_mask = matrix_bd0_24_end_mask_0, x = matrix_bd_24_cast_fp16)[name = string("matrix_bd0_24_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2778_cast_fp16 = add(x = matrix_ac_24_cast_fp16, y = matrix_bd0_24_cast_fp16)[name = string("op_2778_cast_fp16")];
            fp16 _inversed_scores_24_y_0_to_fp16 = const()[name = string("_inversed_scores_24_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_24_cast_fp16 = mul(x = var_2778_cast_fp16, y = _inversed_scores_24_y_0_to_fp16)[name = string("_inversed_scores_24_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_24_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_24_cast_fp16, cond = mask0_4)[name = string("scores0_24_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2784_cast_fp16 = softmax(axis = var_56, x = scores0_24_cast_fp16)[name = string("op_2784_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_143_cast_fp16 = select(a = var_30_to_fp16, b = var_2784_cast_fp16, cond = mask0_4)[name = string("input0_143_cast_fp16")];
            bool x2_24_transpose_x_0 = const()[name = string("x2_24_transpose_x_0"), val = bool(false)];
            bool x2_24_transpose_y_0 = const()[name = string("x2_24_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_26_cast_fp16 = transpose(perm = value_26_perm_0, x = v_24_cast_fp16)[name = string("transpose_257")];
            tensor<fp16, [1, 8, 4, 128]> x2_24_cast_fp16 = matmul(transpose_x = x2_24_transpose_x_0, transpose_y = x2_24_transpose_y_0, x = input0_143_cast_fp16, y = value_26_cast_fp16)[name = string("x2_24_cast_fp16")];
            tensor<int32, [4]> var_2788_perm_0 = const()[name = string("op_2788_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_2788_cast_fp16 = transpose(perm = var_2788_perm_0, x = x2_24_cast_fp16)[name = string("transpose_256")];
            tensor<fp16, [1, 4, 1024]> input1_72_cast_fp16 = reshape(shape = var_2789, x = var_2788_cast_fp16)[name = string("input1_72_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283428736))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284477376))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input1_72_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_145_cast_fp16 = add(x = input_145_cast_fp16, y = linear_106_cast_fp16)[name = string("input0_145_cast_fp16")];
            tensor<int32, [1]> x_235_axes_0 = const()[name = string("x_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284477952)))];
            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284480064)))];
            tensor<fp16, [1, 4, 1024]> x_235_cast_fp16 = layer_norm(axes = x_235_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input0_145_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<int32, [3]> input_149_perm_0 = const()[name = string("input_149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_147_pad_type_0 = const()[name = string("input0_147_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_147_strides_0 = const()[name = string("input0_147_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_147_pad_0 = const()[name = string("input0_147_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_147_dilations_0 = const()[name = string("input0_147_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_147_groups_0 = const()[name = string("input0_147_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284482176))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286579392))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_149_cast_fp16 = transpose(perm = input_149_perm_0, x = x_235_cast_fp16)[name = string("transpose_255")];
            tensor<fp16, [1, 2048, 4]> input0_147_cast_fp16 = conv(dilations = input0_147_dilations_0, groups = input0_147_groups_0, pad = input0_147_pad_0, pad_type = input0_147_pad_type_0, strides = input0_147_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_149_cast_fp16)[name = string("input0_147_cast_fp16")];
            int32 x_237_split_num_splits_0 = const()[name = string("x_237_split_num_splits_0"), val = int32(2)];
            int32 x_237_split_axis_0 = const()[name = string("x_237_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_237_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_237_split_cast_fp16_1 = split(axis = x_237_split_axis_0, num_splits = x_237_split_num_splits_0, x = input0_147_cast_fp16)[name = string("x_237_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_237_split_1_sigmoid_cast_fp16 = sigmoid(x = x_237_split_cast_fp16_1)[name = string("x_237_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_237_cast_fp16 = mul(x = x_237_split_cast_fp16_0, y = x_237_split_1_sigmoid_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_149_cast_fp16 = select(a = var_30_to_fp16, b = x_237_cast_fp16, cond = var_570)[name = string("input0_149_cast_fp16")];
            bool new_x0_24_interleave_0 = const()[name = string("new_x0_24_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_24_cast_fp16 = concat(axis = var_56, interleave = new_x0_24_interleave_0, values = (cache22_1_cast_fp16, input0_149_cast_fp16))[name = string("new_x0_24_cast_fp16")];
            tensor<int32, [3]> var_2827_begin_0 = const()[name = string("op_2827_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2827_end_0 = const()[name = string("op_2827_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_2827_end_mask_0 = const()[name = string("op_2827_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_2827_cast_fp16 = slice_by_index(begin = var_2827_begin_0, end = var_2827_end_0, end_mask = var_2827_end_mask_0, x = new_x0_24_cast_fp16)[name = string("op_2827_cast_fp16")];
            string x_239_pad_type_0 = const()[name = string("x_239_pad_type_0"), val = string("valid")];
            int32 x_239_groups_0 = const()[name = string("x_239_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_239_strides_0 = const()[name = string("x_239_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_239_pad_0 = const()[name = string("x_239_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_239_dilations_0 = const()[name = string("x_239_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286579968))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286589248))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_239_cast_fp16 = conv(dilations = x_239_dilations_0, groups = x_239_groups_0, pad = x_239_pad_0, pad_type = x_239_pad_type_0, strides = x_239_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_24_cast_fp16)[name = string("x_239_cast_fp16")];
            tensor<int32, [3]> input1_74_perm_0 = const()[name = string("input1_74_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_241_axes_0 = const()[name = string("x_241_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286589824)))];
            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286591936)))];
            tensor<fp16, [1, 4, 1024]> input1_74_cast_fp16 = transpose(perm = input1_74_perm_0, x = x_239_cast_fp16)[name = string("transpose_254")];
            tensor<fp16, [1, 4, 1024]> x_241_cast_fp16 = layer_norm(axes = x_241_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input1_74_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<int32, [3]> input2_48_perm_0 = const()[name = string("input2_48_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_48_cast_fp16 = transpose(perm = input2_48_perm_0, x = x_241_cast_fp16)[name = string("transpose_253")];
            tensor<fp16, [1, 1024, 4]> var_2842_cast_fp16 = silu(x = input2_48_cast_fp16)[name = string("op_2842_cast_fp16")];
            string x_243_pad_type_0 = const()[name = string("x_243_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_243_strides_0 = const()[name = string("x_243_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_243_dilations_0 = const()[name = string("x_243_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_243_groups_0 = const()[name = string("x_243_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286594048))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287642688))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_243_cast_fp16 = conv(dilations = x_243_dilations_0, groups = x_243_groups_0, pad = x_243_pad_0, pad_type = x_243_pad_type_0, strides = x_243_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_2842_cast_fp16)[name = string("x_243_cast_fp16")];
            tensor<int32, [3]> input3_26_perm_0 = const()[name = string("input3_26_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_26_cast_fp16 = transpose(perm = input3_26_perm_0, x = x_243_cast_fp16)[name = string("transpose_252")];
            tensor<fp16, [1, 4, 1024]> input1_76_cast_fp16 = add(x = input0_145_cast_fp16, y = input3_26_cast_fp16)[name = string("input1_76_cast_fp16")];
            tensor<int32, [1]> input0_151_axes_0 = const()[name = string("input0_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287643264)))];
            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287645376)))];
            tensor<fp16, [1, 4, 1024]> input0_151_cast_fp16 = layer_norm(axes = input0_151_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input1_76_cast_fp16)[name = string("input0_151_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287647488))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291841856))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_151_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2863_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("op_2863_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291842432))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296036800))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = var_2863_cast_fp16)[name = string("linear_108_cast_fp16")];
            fp16 var_2868_to_fp16 = const()[name = string("op_2868_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2869_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2868_to_fp16)[name = string("op_2869_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_50_cast_fp16 = add(x = input1_76_cast_fp16, y = var_2869_cast_fp16)[name = string("input2_50_cast_fp16")];
            tensor<int32, [1]> input0_153_axes_0 = const()[name = string("input0_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296037376)))];
            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296039488)))];
            tensor<fp16, [1, 4, 1024]> input0_153_cast_fp16 = layer_norm(axes = input0_153_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input2_50_cast_fp16)[name = string("input0_153_cast_fp16")];
            tensor<int32, [4]> cache23_1_begin_0 = const()[name = string("cache23_1_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> cache23_1_end_0 = const()[name = string("cache23_1_end_0"), val = tensor<int32, [4]>([13, 1, 56, 1024])];
            tensor<bool, [4]> cache23_1_end_mask_0 = const()[name = string("cache23_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache23_1_squeeze_mask_0 = const()[name = string("cache23_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache23_1_cast_fp16 = slice_by_index(begin = cache23_1_begin_0, end = cache23_1_end_0, end_mask = cache23_1_end_mask_0, squeeze_mask = cache23_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache23_1_cast_fp16")];
            tensor<int32, [4]> cache24_1_begin_0 = const()[name = string("cache24_1_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> cache24_1_end_0 = const()[name = string("cache24_1_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
            tensor<bool, [4]> cache24_1_end_mask_0 = const()[name = string("cache24_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache24_1_squeeze_mask_0 = const()[name = string("cache24_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache24_1_cast_fp16 = slice_by_index(begin = cache24_1_begin_0, end = cache24_1_end_0, end_mask = cache24_1_end_mask_0, squeeze_mask = cache24_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache24_1_cast_fp16")];
            tensor<int32, [1]> input_153_axes_0 = const()[name = string("input_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296041600)))];
            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296043712)))];
            tensor<fp16, [1, 4, 1024]> input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input0_153_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296045824))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300240192))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_2898_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2898_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300240768))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304435136))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = var_2898_cast_fp16)[name = string("linear_110_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_2904_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2903_to_fp16)[name = string("op_2904_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_157_cast_fp16 = add(x = input0_153_cast_fp16, y = var_2904_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<int32, [1]> key_26_axes_0 = const()[name = string("key_26_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304435712)))];
            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304437824)))];
            tensor<fp16, [1, 4, 1024]> key_26_cast_fp16 = layer_norm(axes = key_26_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_157_cast_fp16)[name = string("key_26_cast_fp16")];
            bool input_159_interleave_0 = const()[name = string("input_159_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_159_cast_fp16 = concat(axis = var_64, interleave = input_159_interleave_0, values = (cache23_1_cast_fp16, key_26_cast_fp16))[name = string("input_159_cast_fp16")];
            tensor<int32, [3]> var_2926_begin_0 = const()[name = string("op_2926_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_2926_end_0 = const()[name = string("op_2926_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_2926_end_mask_0 = const()[name = string("op_2926_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_2926_cast_fp16 = slice_by_index(begin = var_2926_begin_0, end = var_2926_end_0, end_mask = var_2926_end_mask_0, x = cache23_1_cast_fp16)[name = string("op_2926_cast_fp16")];
            bool var_2932_interleave_0 = const()[name = string("op_2932_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_2932_cast_fp16 = concat(axis = var_64, interleave = var_2932_interleave_0, values = (var_2926_cast_fp16, key_26_cast_fp16))[name = string("op_2932_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304439936))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305488576))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_26_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_2936 = const()[name = string("op_2936"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_26_cast_fp16 = reshape(shape = var_2936, x = linear_111_cast_fp16)[name = string("q_26_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305489152))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306537792))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_159_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_26_cast_fp16 = reshape(shape = var_2940, x = linear_112_cast_fp16)[name = string("k_26_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306538368))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307587008))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_159_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2944 = const()[name = string("op_2944"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_26_cast_fp16 = reshape(shape = var_2944, x = linear_113_cast_fp16)[name = string("v_26_cast_fp16")];
            tensor<int32, [4]> value_28_perm_0 = const()[name = string("value_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307587584)))];
            tensor<fp16, [1, 4, 8, 128]> var_2956_cast_fp16 = add(x = q_26_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_2956_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307589696)))];
            tensor<fp16, [1, 4, 8, 128]> var_2958_cast_fp16 = add(x = q_26_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_2958_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_26_perm_0 = const()[name = string("q_with_bias_v_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_251_transpose_x_0 = const()[name = string("x_251_transpose_x_0"), val = bool(false)];
            bool x_251_transpose_y_0 = const()[name = string("x_251_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_2960_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307591808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307713728))))[name = string("op_2960_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_26_cast_fp16 = transpose(perm = q_with_bias_v_26_perm_0, x = var_2958_cast_fp16)[name = string("transpose_251")];
            tensor<fp16, [1, 8, 4, 119]> x_251_cast_fp16 = matmul(transpose_x = x_251_transpose_x_0, transpose_y = x_251_transpose_y_0, x = q_with_bias_v_26_cast_fp16, y = op_2960_to_fp16_palettized)[name = string("x_251_cast_fp16")];
            tensor<int32, [8]> x0_28_pad_0 = const()[name = string("x0_28_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_28_mode_0 = const()[name = string("x0_28_mode_0"), val = string("constant")];
            fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_28_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x0_28_mode_0, pad = x0_28_pad_0, x = x_251_cast_fp16)[name = string("x0_28_cast_fp16")];
            tensor<int32, [4]> var_2968 = const()[name = string("op_2968"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_26_cast_fp16 = reshape(shape = var_2968, x = x0_28_cast_fp16)[name = string("x1_26_cast_fp16")];
            tensor<int32, [4]> var_2972_begin_0 = const()[name = string("op_2972_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2972_end_0 = const()[name = string("op_2972_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_2972_end_mask_0 = const()[name = string("op_2972_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_2972_cast_fp16 = slice_by_index(begin = var_2972_begin_0, end = var_2972_end_0, end_mask = var_2972_end_mask_0, x = x1_26_cast_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2973 = const()[name = string("op_2973"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_26_cast_fp16 = reshape(shape = var_2973, x = var_2972_cast_fp16)[name = string("matrix_bd_26_cast_fp16")];
            bool matrix_ac_26_transpose_x_0 = const()[name = string("matrix_ac_26_transpose_x_0"), val = bool(false)];
            bool matrix_ac_26_transpose_y_0 = const()[name = string("matrix_ac_26_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_26_cast_fp16)[name = string("transpose_249")];
            tensor<fp16, [1, 8, 4, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_2956_cast_fp16)[name = string("transpose_250")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_26_cast_fp16 = matmul(transpose_x = matrix_ac_26_transpose_x_0, transpose_y = matrix_ac_26_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_26_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_26_begin_0 = const()[name = string("matrix_bd0_26_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_26_end_0 = const()[name = string("matrix_bd0_26_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_26_end_mask_0 = const()[name = string("matrix_bd0_26_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_26_cast_fp16 = slice_by_index(begin = matrix_bd0_26_begin_0, end = matrix_bd0_26_end_0, end_mask = matrix_bd0_26_end_mask_0, x = matrix_bd_26_cast_fp16)[name = string("matrix_bd0_26_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2982_cast_fp16 = add(x = matrix_ac_26_cast_fp16, y = matrix_bd0_26_cast_fp16)[name = string("op_2982_cast_fp16")];
            fp16 _inversed_scores_26_y_0_to_fp16 = const()[name = string("_inversed_scores_26_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_26_cast_fp16 = mul(x = var_2982_cast_fp16, y = _inversed_scores_26_y_0_to_fp16)[name = string("_inversed_scores_26_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_26_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_26_cast_fp16, cond = mask0_4)[name = string("scores0_26_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_2988_cast_fp16 = softmax(axis = var_56, x = scores0_26_cast_fp16)[name = string("op_2988_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_155_cast_fp16 = select(a = var_30_to_fp16, b = var_2988_cast_fp16, cond = mask0_4)[name = string("input0_155_cast_fp16")];
            bool x2_26_transpose_x_0 = const()[name = string("x2_26_transpose_x_0"), val = bool(false)];
            bool x2_26_transpose_y_0 = const()[name = string("x2_26_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_28_cast_fp16 = transpose(perm = value_28_perm_0, x = v_26_cast_fp16)[name = string("transpose_248")];
            tensor<fp16, [1, 8, 4, 128]> x2_26_cast_fp16 = matmul(transpose_x = x2_26_transpose_x_0, transpose_y = x2_26_transpose_y_0, x = input0_155_cast_fp16, y = value_28_cast_fp16)[name = string("x2_26_cast_fp16")];
            tensor<int32, [4]> var_2992_perm_0 = const()[name = string("op_2992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2993 = const()[name = string("op_2993"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_2992_cast_fp16 = transpose(perm = var_2992_perm_0, x = x2_26_cast_fp16)[name = string("transpose_247")];
            tensor<fp16, [1, 4, 1024]> input1_78_cast_fp16 = reshape(shape = var_2993, x = var_2992_cast_fp16)[name = string("input1_78_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307714304))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308762944))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input1_78_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_157_cast_fp16 = add(x = input_157_cast_fp16, y = linear_115_cast_fp16)[name = string("input0_157_cast_fp16")];
            tensor<int32, [1]> x_255_axes_0 = const()[name = string("x_255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308763520)))];
            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308765632)))];
            tensor<fp16, [1, 4, 1024]> x_255_cast_fp16 = layer_norm(axes = x_255_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input0_157_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_159_pad_type_0 = const()[name = string("input0_159_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_159_strides_0 = const()[name = string("input0_159_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_159_pad_0 = const()[name = string("input0_159_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_159_dilations_0 = const()[name = string("input0_159_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_159_groups_0 = const()[name = string("input0_159_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308767744))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310864960))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_255_cast_fp16)[name = string("transpose_246")];
            tensor<fp16, [1, 2048, 4]> input0_159_cast_fp16 = conv(dilations = input0_159_dilations_0, groups = input0_159_groups_0, pad = input0_159_pad_0, pad_type = input0_159_pad_type_0, strides = input0_159_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_161_cast_fp16)[name = string("input0_159_cast_fp16")];
            int32 x_257_split_num_splits_0 = const()[name = string("x_257_split_num_splits_0"), val = int32(2)];
            int32 x_257_split_axis_0 = const()[name = string("x_257_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_257_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_257_split_cast_fp16_1 = split(axis = x_257_split_axis_0, num_splits = x_257_split_num_splits_0, x = input0_159_cast_fp16)[name = string("x_257_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_257_split_1_sigmoid_cast_fp16 = sigmoid(x = x_257_split_cast_fp16_1)[name = string("x_257_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_257_cast_fp16 = mul(x = x_257_split_cast_fp16_0, y = x_257_split_1_sigmoid_cast_fp16)[name = string("x_257_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_161_cast_fp16 = select(a = var_30_to_fp16, b = x_257_cast_fp16, cond = var_570)[name = string("input0_161_cast_fp16")];
            bool new_x0_26_interleave_0 = const()[name = string("new_x0_26_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_26_cast_fp16 = concat(axis = var_56, interleave = new_x0_26_interleave_0, values = (cache24_1_cast_fp16, input0_161_cast_fp16))[name = string("new_x0_26_cast_fp16")];
            tensor<int32, [3]> var_3031_begin_0 = const()[name = string("op_3031_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3031_end_0 = const()[name = string("op_3031_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_3031_end_mask_0 = const()[name = string("op_3031_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_3031_cast_fp16 = slice_by_index(begin = var_3031_begin_0, end = var_3031_end_0, end_mask = var_3031_end_mask_0, x = new_x0_26_cast_fp16)[name = string("op_3031_cast_fp16")];
            string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")];
            int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310865536))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310874816))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_26_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<int32, [3]> input1_80_perm_0 = const()[name = string("input1_80_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_261_axes_0 = const()[name = string("x_261_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310875392)))];
            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310877504)))];
            tensor<fp16, [1, 4, 1024]> input1_80_cast_fp16 = transpose(perm = input1_80_perm_0, x = x_259_cast_fp16)[name = string("transpose_245")];
            tensor<fp16, [1, 4, 1024]> x_261_cast_fp16 = layer_norm(axes = x_261_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input1_80_cast_fp16)[name = string("x_261_cast_fp16")];
            tensor<int32, [3]> input2_52_perm_0 = const()[name = string("input2_52_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_52_cast_fp16 = transpose(perm = input2_52_perm_0, x = x_261_cast_fp16)[name = string("transpose_244")];
            tensor<fp16, [1, 1024, 4]> var_3046_cast_fp16 = silu(x = input2_52_cast_fp16)[name = string("op_3046_cast_fp16")];
            string x_263_pad_type_0 = const()[name = string("x_263_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_263_strides_0 = const()[name = string("x_263_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_263_pad_0 = const()[name = string("x_263_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_263_dilations_0 = const()[name = string("x_263_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_263_groups_0 = const()[name = string("x_263_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310879616))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311928256))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_263_cast_fp16 = conv(dilations = x_263_dilations_0, groups = x_263_groups_0, pad = x_263_pad_0, pad_type = x_263_pad_type_0, strides = x_263_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_3046_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<int32, [3]> input3_28_perm_0 = const()[name = string("input3_28_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_28_cast_fp16 = transpose(perm = input3_28_perm_0, x = x_263_cast_fp16)[name = string("transpose_243")];
            tensor<fp16, [1, 4, 1024]> input1_82_cast_fp16 = add(x = input0_157_cast_fp16, y = input3_28_cast_fp16)[name = string("input1_82_cast_fp16")];
            tensor<int32, [1]> input0_163_axes_0 = const()[name = string("input0_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311928832)))];
            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311930944)))];
            tensor<fp16, [1, 4, 1024]> input0_163_cast_fp16 = layer_norm(axes = input0_163_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input1_82_cast_fp16)[name = string("input0_163_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311933056))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316127424))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_163_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3067_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316128000))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320322368))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = var_3067_cast_fp16)[name = string("linear_117_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3073_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_54_cast_fp16 = add(x = input1_82_cast_fp16, y = var_3073_cast_fp16)[name = string("input2_54_cast_fp16")];
            tensor<int32, [1]> input0_165_axes_0 = const()[name = string("input0_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320322944)))];
            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320325056)))];
            tensor<fp16, [1, 4, 1024]> input0_165_cast_fp16 = layer_norm(axes = input0_165_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input2_54_cast_fp16)[name = string("input0_165_cast_fp16")];
            tensor<int32, [4]> cache25_1_begin_0 = const()[name = string("cache25_1_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> cache25_1_end_0 = const()[name = string("cache25_1_end_0"), val = tensor<int32, [4]>([14, 1, 56, 1024])];
            tensor<bool, [4]> cache25_1_end_mask_0 = const()[name = string("cache25_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache25_1_squeeze_mask_0 = const()[name = string("cache25_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache25_1_cast_fp16 = slice_by_index(begin = cache25_1_begin_0, end = cache25_1_end_0, end_mask = cache25_1_end_mask_0, squeeze_mask = cache25_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache25_1_cast_fp16")];
            tensor<int32, [4]> cache26_1_begin_0 = const()[name = string("cache26_1_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> cache26_1_end_0 = const()[name = string("cache26_1_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
            tensor<bool, [4]> cache26_1_end_mask_0 = const()[name = string("cache26_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache26_1_squeeze_mask_0 = const()[name = string("cache26_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache26_1_cast_fp16 = slice_by_index(begin = cache26_1_begin_0, end = cache26_1_end_0, end_mask = cache26_1_end_mask_0, squeeze_mask = cache26_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache26_1_cast_fp16")];
            tensor<int32, [1]> input_165_axes_0 = const()[name = string("input_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320327168)))];
            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320329280)))];
            tensor<fp16, [1, 4, 1024]> input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input0_165_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320331392))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324525760))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_165_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3102_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("op_3102_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324526336))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328720704))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = var_3102_cast_fp16)[name = string("linear_119_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3108_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_169_cast_fp16 = add(x = input0_165_cast_fp16, y = var_3108_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<int32, [1]> key_28_axes_0 = const()[name = string("key_28_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328721280)))];
            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328723392)))];
            tensor<fp16, [1, 4, 1024]> key_28_cast_fp16 = layer_norm(axes = key_28_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_169_cast_fp16)[name = string("key_28_cast_fp16")];
            bool input_171_interleave_0 = const()[name = string("input_171_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_171_cast_fp16 = concat(axis = var_64, interleave = input_171_interleave_0, values = (cache25_1_cast_fp16, key_28_cast_fp16))[name = string("input_171_cast_fp16")];
            tensor<int32, [3]> var_3130_begin_0 = const()[name = string("op_3130_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3130_end_0 = const()[name = string("op_3130_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_3130_end_mask_0 = const()[name = string("op_3130_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_3130_cast_fp16 = slice_by_index(begin = var_3130_begin_0, end = var_3130_end_0, end_mask = var_3130_end_mask_0, x = cache25_1_cast_fp16)[name = string("op_3130_cast_fp16")];
            bool var_3136_interleave_0 = const()[name = string("op_3136_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_3136_cast_fp16 = concat(axis = var_64, interleave = var_3136_interleave_0, values = (var_3130_cast_fp16, key_28_cast_fp16))[name = string("op_3136_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328725504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329774144))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_28_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_3140 = const()[name = string("op_3140"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_28_cast_fp16 = reshape(shape = var_3140, x = linear_120_cast_fp16)[name = string("q_28_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329774720))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330823360))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_3144 = const()[name = string("op_3144"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_28_cast_fp16 = reshape(shape = var_3144, x = linear_121_cast_fp16)[name = string("k_28_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330823936))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331872576))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<int32, [4]> var_3148 = const()[name = string("op_3148"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_28_cast_fp16 = reshape(shape = var_3148, x = linear_122_cast_fp16)[name = string("v_28_cast_fp16")];
            tensor<int32, [4]> value_30_perm_0 = const()[name = string("value_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331873152)))];
            tensor<fp16, [1, 4, 8, 128]> var_3160_cast_fp16 = add(x = q_28_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331875264)))];
            tensor<fp16, [1, 4, 8, 128]> var_3162_cast_fp16 = add(x = q_28_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3162_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_28_perm_0 = const()[name = string("q_with_bias_v_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_271_transpose_x_0 = const()[name = string("x_271_transpose_x_0"), val = bool(false)];
            bool x_271_transpose_y_0 = const()[name = string("x_271_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_3164_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331877376))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331999296))))[name = string("op_3164_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_28_cast_fp16 = transpose(perm = q_with_bias_v_28_perm_0, x = var_3162_cast_fp16)[name = string("transpose_242")];
            tensor<fp16, [1, 8, 4, 119]> x_271_cast_fp16 = matmul(transpose_x = x_271_transpose_x_0, transpose_y = x_271_transpose_y_0, x = q_with_bias_v_28_cast_fp16, y = op_3164_to_fp16_palettized)[name = string("x_271_cast_fp16")];
            tensor<int32, [8]> x0_30_pad_0 = const()[name = string("x0_30_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_30_mode_0 = const()[name = string("x0_30_mode_0"), val = string("constant")];
            fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_30_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x0_30_mode_0, pad = x0_30_pad_0, x = x_271_cast_fp16)[name = string("x0_30_cast_fp16")];
            tensor<int32, [4]> var_3172 = const()[name = string("op_3172"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_28_cast_fp16 = reshape(shape = var_3172, x = x0_30_cast_fp16)[name = string("x1_28_cast_fp16")];
            tensor<int32, [4]> var_3176_begin_0 = const()[name = string("op_3176_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3176_end_0 = const()[name = string("op_3176_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_3176_end_mask_0 = const()[name = string("op_3176_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_3176_cast_fp16 = slice_by_index(begin = var_3176_begin_0, end = var_3176_end_0, end_mask = var_3176_end_mask_0, x = x1_28_cast_fp16)[name = string("op_3176_cast_fp16")];
            tensor<int32, [4]> var_3177 = const()[name = string("op_3177"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_28_cast_fp16 = reshape(shape = var_3177, x = var_3176_cast_fp16)[name = string("matrix_bd_28_cast_fp16")];
            bool matrix_ac_28_transpose_x_0 = const()[name = string("matrix_ac_28_transpose_x_0"), val = bool(false)];
            bool matrix_ac_28_transpose_y_0 = const()[name = string("matrix_ac_28_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_28_cast_fp16)[name = string("transpose_240")];
            tensor<fp16, [1, 8, 4, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3160_cast_fp16)[name = string("transpose_241")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_28_cast_fp16 = matmul(transpose_x = matrix_ac_28_transpose_x_0, transpose_y = matrix_ac_28_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_28_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_28_begin_0 = const()[name = string("matrix_bd0_28_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_28_end_0 = const()[name = string("matrix_bd0_28_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_28_end_mask_0 = const()[name = string("matrix_bd0_28_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_28_cast_fp16 = slice_by_index(begin = matrix_bd0_28_begin_0, end = matrix_bd0_28_end_0, end_mask = matrix_bd0_28_end_mask_0, x = matrix_bd_28_cast_fp16)[name = string("matrix_bd0_28_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3186_cast_fp16 = add(x = matrix_ac_28_cast_fp16, y = matrix_bd0_28_cast_fp16)[name = string("op_3186_cast_fp16")];
            fp16 _inversed_scores_28_y_0_to_fp16 = const()[name = string("_inversed_scores_28_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_28_cast_fp16 = mul(x = var_3186_cast_fp16, y = _inversed_scores_28_y_0_to_fp16)[name = string("_inversed_scores_28_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_28_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_28_cast_fp16, cond = mask0_4)[name = string("scores0_28_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3192_cast_fp16 = softmax(axis = var_56, x = scores0_28_cast_fp16)[name = string("op_3192_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_167_cast_fp16 = select(a = var_30_to_fp16, b = var_3192_cast_fp16, cond = mask0_4)[name = string("input0_167_cast_fp16")];
            bool x2_28_transpose_x_0 = const()[name = string("x2_28_transpose_x_0"), val = bool(false)];
            bool x2_28_transpose_y_0 = const()[name = string("x2_28_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_30_cast_fp16 = transpose(perm = value_30_perm_0, x = v_28_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 8, 4, 128]> x2_28_cast_fp16 = matmul(transpose_x = x2_28_transpose_x_0, transpose_y = x2_28_transpose_y_0, x = input0_167_cast_fp16, y = value_30_cast_fp16)[name = string("x2_28_cast_fp16")];
            tensor<int32, [4]> var_3196_perm_0 = const()[name = string("op_3196_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3197 = const()[name = string("op_3197"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_3196_cast_fp16 = transpose(perm = var_3196_perm_0, x = x2_28_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 4, 1024]> input1_84_cast_fp16 = reshape(shape = var_3197, x = var_3196_cast_fp16)[name = string("input1_84_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331999872))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333048512))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input1_84_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_169_cast_fp16 = add(x = input_169_cast_fp16, y = linear_124_cast_fp16)[name = string("input0_169_cast_fp16")];
            tensor<int32, [1]> x_275_axes_0 = const()[name = string("x_275_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333049088)))];
            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333051200)))];
            tensor<fp16, [1, 4, 1024]> x_275_cast_fp16 = layer_norm(axes = x_275_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input0_169_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<int32, [3]> input_173_perm_0 = const()[name = string("input_173_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_171_pad_type_0 = const()[name = string("input0_171_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_171_strides_0 = const()[name = string("input0_171_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_171_pad_0 = const()[name = string("input0_171_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_171_dilations_0 = const()[name = string("input0_171_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_171_groups_0 = const()[name = string("input0_171_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333053312))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335150528))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_173_cast_fp16 = transpose(perm = input_173_perm_0, x = x_275_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2048, 4]> input0_171_cast_fp16 = conv(dilations = input0_171_dilations_0, groups = input0_171_groups_0, pad = input0_171_pad_0, pad_type = input0_171_pad_type_0, strides = input0_171_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("input0_171_cast_fp16")];
            int32 x_277_split_num_splits_0 = const()[name = string("x_277_split_num_splits_0"), val = int32(2)];
            int32 x_277_split_axis_0 = const()[name = string("x_277_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_277_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_277_split_cast_fp16_1 = split(axis = x_277_split_axis_0, num_splits = x_277_split_num_splits_0, x = input0_171_cast_fp16)[name = string("x_277_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_277_split_1_sigmoid_cast_fp16 = sigmoid(x = x_277_split_cast_fp16_1)[name = string("x_277_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_277_cast_fp16 = mul(x = x_277_split_cast_fp16_0, y = x_277_split_1_sigmoid_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_173_cast_fp16 = select(a = var_30_to_fp16, b = x_277_cast_fp16, cond = var_570)[name = string("input0_173_cast_fp16")];
            bool new_x0_28_interleave_0 = const()[name = string("new_x0_28_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_28_cast_fp16 = concat(axis = var_56, interleave = new_x0_28_interleave_0, values = (cache26_1_cast_fp16, input0_173_cast_fp16))[name = string("new_x0_28_cast_fp16")];
            tensor<int32, [3]> var_3235_begin_0 = const()[name = string("op_3235_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3235_end_0 = const()[name = string("op_3235_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_3235_end_mask_0 = const()[name = string("op_3235_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_3235_cast_fp16 = slice_by_index(begin = var_3235_begin_0, end = var_3235_end_0, end_mask = var_3235_end_mask_0, x = new_x0_28_cast_fp16)[name = string("op_3235_cast_fp16")];
            string x_279_pad_type_0 = const()[name = string("x_279_pad_type_0"), val = string("valid")];
            int32 x_279_groups_0 = const()[name = string("x_279_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_279_strides_0 = const()[name = string("x_279_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_279_pad_0 = const()[name = string("x_279_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_279_dilations_0 = const()[name = string("x_279_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335151104))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335160384))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_279_cast_fp16 = conv(dilations = x_279_dilations_0, groups = x_279_groups_0, pad = x_279_pad_0, pad_type = x_279_pad_type_0, strides = x_279_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_28_cast_fp16)[name = string("x_279_cast_fp16")];
            tensor<int32, [3]> input1_86_perm_0 = const()[name = string("input1_86_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_281_axes_0 = const()[name = string("x_281_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335160960)))];
            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335163072)))];
            tensor<fp16, [1, 4, 1024]> input1_86_cast_fp16 = transpose(perm = input1_86_perm_0, x = x_279_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [1, 4, 1024]> x_281_cast_fp16 = layer_norm(axes = x_281_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input1_86_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<int32, [3]> input2_56_perm_0 = const()[name = string("input2_56_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_56_cast_fp16 = transpose(perm = input2_56_perm_0, x = x_281_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [1, 1024, 4]> var_3250_cast_fp16 = silu(x = input2_56_cast_fp16)[name = string("op_3250_cast_fp16")];
            string x_283_pad_type_0 = const()[name = string("x_283_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_283_strides_0 = const()[name = string("x_283_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_283_pad_0 = const()[name = string("x_283_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_283_dilations_0 = const()[name = string("x_283_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_283_groups_0 = const()[name = string("x_283_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335165184))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336213824))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_283_cast_fp16 = conv(dilations = x_283_dilations_0, groups = x_283_groups_0, pad = x_283_pad_0, pad_type = x_283_pad_type_0, strides = x_283_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_3250_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<int32, [3]> input3_30_perm_0 = const()[name = string("input3_30_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_30_cast_fp16 = transpose(perm = input3_30_perm_0, x = x_283_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [1, 4, 1024]> input1_88_cast_fp16 = add(x = input0_169_cast_fp16, y = input3_30_cast_fp16)[name = string("input1_88_cast_fp16")];
            tensor<int32, [1]> input0_175_axes_0 = const()[name = string("input0_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336214400)))];
            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336216512)))];
            tensor<fp16, [1, 4, 1024]> input0_175_cast_fp16 = layer_norm(axes = input0_175_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input1_88_cast_fp16)[name = string("input0_175_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336218624))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340412992))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_175_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3271_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340413568))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344607936))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = var_3271_cast_fp16)[name = string("linear_126_cast_fp16")];
            fp16 var_3276_to_fp16 = const()[name = string("op_3276_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3277_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3276_to_fp16)[name = string("op_3277_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_58_cast_fp16 = add(x = input1_88_cast_fp16, y = var_3277_cast_fp16)[name = string("input2_58_cast_fp16")];
            tensor<int32, [1]> input0_177_axes_0 = const()[name = string("input0_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344608512)))];
            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344610624)))];
            tensor<fp16, [1, 4, 1024]> input0_177_cast_fp16 = layer_norm(axes = input0_177_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input2_58_cast_fp16)[name = string("input0_177_cast_fp16")];
            tensor<int32, [4]> cache27_1_begin_0 = const()[name = string("cache27_1_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> cache27_1_end_0 = const()[name = string("cache27_1_end_0"), val = tensor<int32, [4]>([15, 1, 56, 1024])];
            tensor<bool, [4]> cache27_1_end_mask_0 = const()[name = string("cache27_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache27_1_squeeze_mask_0 = const()[name = string("cache27_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache27_1_cast_fp16 = slice_by_index(begin = cache27_1_begin_0, end = cache27_1_end_0, end_mask = cache27_1_end_mask_0, squeeze_mask = cache27_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache27_1_cast_fp16")];
            tensor<int32, [4]> cache28_1_begin_0 = const()[name = string("cache28_1_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> cache28_1_end_0 = const()[name = string("cache28_1_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
            tensor<bool, [4]> cache28_1_end_mask_0 = const()[name = string("cache28_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache28_1_squeeze_mask_0 = const()[name = string("cache28_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache28_1_cast_fp16 = slice_by_index(begin = cache28_1_begin_0, end = cache28_1_end_0, end_mask = cache28_1_end_mask_0, squeeze_mask = cache28_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache28_1_cast_fp16")];
            tensor<int32, [1]> input_177_axes_0 = const()[name = string("input_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344612736)))];
            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344614848)))];
            tensor<fp16, [1, 4, 1024]> input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input0_177_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344616960))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348811328))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3306_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("op_3306_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348811904))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353006272))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = var_3306_cast_fp16)[name = string("linear_128_cast_fp16")];
            fp16 var_3311_to_fp16 = const()[name = string("op_3311_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3312_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3311_to_fp16)[name = string("op_3312_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_181_cast_fp16 = add(x = input0_177_cast_fp16, y = var_3312_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<int32, [1]> key_30_axes_0 = const()[name = string("key_30_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353006848)))];
            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353008960)))];
            tensor<fp16, [1, 4, 1024]> key_30_cast_fp16 = layer_norm(axes = key_30_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_181_cast_fp16)[name = string("key_30_cast_fp16")];
            bool input_183_interleave_0 = const()[name = string("input_183_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_183_cast_fp16 = concat(axis = var_64, interleave = input_183_interleave_0, values = (cache27_1_cast_fp16, key_30_cast_fp16))[name = string("input_183_cast_fp16")];
            tensor<int32, [3]> var_3334_begin_0 = const()[name = string("op_3334_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3334_end_0 = const()[name = string("op_3334_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_3334_end_mask_0 = const()[name = string("op_3334_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_3334_cast_fp16 = slice_by_index(begin = var_3334_begin_0, end = var_3334_end_0, end_mask = var_3334_end_mask_0, x = cache27_1_cast_fp16)[name = string("op_3334_cast_fp16")];
            bool var_3340_interleave_0 = const()[name = string("op_3340_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_3340_cast_fp16 = concat(axis = var_64, interleave = var_3340_interleave_0, values = (var_3334_cast_fp16, key_30_cast_fp16))[name = string("op_3340_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353011072))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354059712))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_30_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_3344 = const()[name = string("op_3344"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_30_cast_fp16 = reshape(shape = var_3344, x = linear_129_cast_fp16)[name = string("q_30_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354060288))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355108928))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<int32, [4]> var_3348 = const()[name = string("op_3348"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_30_cast_fp16 = reshape(shape = var_3348, x = linear_130_cast_fp16)[name = string("k_30_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355109504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356158144))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<int32, [4]> var_3352 = const()[name = string("op_3352"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_30_cast_fp16 = reshape(shape = var_3352, x = linear_131_cast_fp16)[name = string("v_30_cast_fp16")];
            tensor<int32, [4]> value_32_perm_0 = const()[name = string("value_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356158720)))];
            tensor<fp16, [1, 4, 8, 128]> var_3364_cast_fp16 = add(x = q_30_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3364_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356160832)))];
            tensor<fp16, [1, 4, 8, 128]> var_3366_cast_fp16 = add(x = q_30_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3366_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_30_perm_0 = const()[name = string("q_with_bias_v_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_291_transpose_x_0 = const()[name = string("x_291_transpose_x_0"), val = bool(false)];
            bool x_291_transpose_y_0 = const()[name = string("x_291_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_3368_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356162944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356284864))))[name = string("op_3368_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_30_cast_fp16 = transpose(perm = q_with_bias_v_30_perm_0, x = var_3366_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [1, 8, 4, 119]> x_291_cast_fp16 = matmul(transpose_x = x_291_transpose_x_0, transpose_y = x_291_transpose_y_0, x = q_with_bias_v_30_cast_fp16, y = op_3368_to_fp16_palettized)[name = string("x_291_cast_fp16")];
            tensor<int32, [8]> x0_32_pad_0 = const()[name = string("x0_32_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_32_mode_0 = const()[name = string("x0_32_mode_0"), val = string("constant")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_32_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x0_32_mode_0, pad = x0_32_pad_0, x = x_291_cast_fp16)[name = string("x0_32_cast_fp16")];
            tensor<int32, [4]> var_3376 = const()[name = string("op_3376"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_30_cast_fp16 = reshape(shape = var_3376, x = x0_32_cast_fp16)[name = string("x1_30_cast_fp16")];
            tensor<int32, [4]> var_3380_begin_0 = const()[name = string("op_3380_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3380_end_0 = const()[name = string("op_3380_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_3380_end_mask_0 = const()[name = string("op_3380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_3380_cast_fp16 = slice_by_index(begin = var_3380_begin_0, end = var_3380_end_0, end_mask = var_3380_end_mask_0, x = x1_30_cast_fp16)[name = string("op_3380_cast_fp16")];
            tensor<int32, [4]> var_3381 = const()[name = string("op_3381"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_30_cast_fp16 = reshape(shape = var_3381, x = var_3380_cast_fp16)[name = string("matrix_bd_30_cast_fp16")];
            bool matrix_ac_30_transpose_x_0 = const()[name = string("matrix_ac_30_transpose_x_0"), val = bool(false)];
            bool matrix_ac_30_transpose_y_0 = const()[name = string("matrix_ac_30_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_30_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 8, 4, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3364_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_30_cast_fp16 = matmul(transpose_x = matrix_ac_30_transpose_x_0, transpose_y = matrix_ac_30_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_30_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_30_begin_0 = const()[name = string("matrix_bd0_30_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_30_end_0 = const()[name = string("matrix_bd0_30_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_30_end_mask_0 = const()[name = string("matrix_bd0_30_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_30_cast_fp16 = slice_by_index(begin = matrix_bd0_30_begin_0, end = matrix_bd0_30_end_0, end_mask = matrix_bd0_30_end_mask_0, x = matrix_bd_30_cast_fp16)[name = string("matrix_bd0_30_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3390_cast_fp16 = add(x = matrix_ac_30_cast_fp16, y = matrix_bd0_30_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 _inversed_scores_30_y_0_to_fp16 = const()[name = string("_inversed_scores_30_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_30_cast_fp16 = mul(x = var_3390_cast_fp16, y = _inversed_scores_30_y_0_to_fp16)[name = string("_inversed_scores_30_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_30_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_30_cast_fp16, cond = mask0_4)[name = string("scores0_30_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3396_cast_fp16 = softmax(axis = var_56, x = scores0_30_cast_fp16)[name = string("op_3396_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_179_cast_fp16 = select(a = var_30_to_fp16, b = var_3396_cast_fp16, cond = mask0_4)[name = string("input0_179_cast_fp16")];
            bool x2_30_transpose_x_0 = const()[name = string("x2_30_transpose_x_0"), val = bool(false)];
            bool x2_30_transpose_y_0 = const()[name = string("x2_30_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_32_cast_fp16 = transpose(perm = value_32_perm_0, x = v_30_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 8, 4, 128]> x2_30_cast_fp16 = matmul(transpose_x = x2_30_transpose_x_0, transpose_y = x2_30_transpose_y_0, x = input0_179_cast_fp16, y = value_32_cast_fp16)[name = string("x2_30_cast_fp16")];
            tensor<int32, [4]> var_3400_perm_0 = const()[name = string("op_3400_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3401 = const()[name = string("op_3401"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_3400_cast_fp16 = transpose(perm = var_3400_perm_0, x = x2_30_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 1024]> input1_90_cast_fp16 = reshape(shape = var_3401, x = var_3400_cast_fp16)[name = string("input1_90_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356285440))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357334080))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input1_90_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_181_cast_fp16 = add(x = input_181_cast_fp16, y = linear_133_cast_fp16)[name = string("input0_181_cast_fp16")];
            tensor<int32, [1]> x_295_axes_0 = const()[name = string("x_295_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357334656)))];
            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357336768)))];
            tensor<fp16, [1, 4, 1024]> x_295_cast_fp16 = layer_norm(axes = x_295_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input0_181_cast_fp16)[name = string("x_295_cast_fp16")];
            tensor<int32, [3]> input_185_perm_0 = const()[name = string("input_185_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_183_pad_type_0 = const()[name = string("input0_183_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_183_strides_0 = const()[name = string("input0_183_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_183_pad_0 = const()[name = string("input0_183_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_183_dilations_0 = const()[name = string("input0_183_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_183_groups_0 = const()[name = string("input0_183_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357338880))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359436096))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_185_cast_fp16 = transpose(perm = input_185_perm_0, x = x_295_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2048, 4]> input0_183_cast_fp16 = conv(dilations = input0_183_dilations_0, groups = input0_183_groups_0, pad = input0_183_pad_0, pad_type = input0_183_pad_type_0, strides = input0_183_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_185_cast_fp16)[name = string("input0_183_cast_fp16")];
            int32 x_297_split_num_splits_0 = const()[name = string("x_297_split_num_splits_0"), val = int32(2)];
            int32 x_297_split_axis_0 = const()[name = string("x_297_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_297_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_297_split_cast_fp16_1 = split(axis = x_297_split_axis_0, num_splits = x_297_split_num_splits_0, x = input0_183_cast_fp16)[name = string("x_297_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_297_split_1_sigmoid_cast_fp16 = sigmoid(x = x_297_split_cast_fp16_1)[name = string("x_297_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_297_cast_fp16 = mul(x = x_297_split_cast_fp16_0, y = x_297_split_1_sigmoid_cast_fp16)[name = string("x_297_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_185_cast_fp16 = select(a = var_30_to_fp16, b = x_297_cast_fp16, cond = var_570)[name = string("input0_185_cast_fp16")];
            bool new_x0_30_interleave_0 = const()[name = string("new_x0_30_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_30_cast_fp16 = concat(axis = var_56, interleave = new_x0_30_interleave_0, values = (cache28_1_cast_fp16, input0_185_cast_fp16))[name = string("new_x0_30_cast_fp16")];
            tensor<int32, [3]> var_3439_begin_0 = const()[name = string("op_3439_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3439_end_0 = const()[name = string("op_3439_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_3439_end_mask_0 = const()[name = string("op_3439_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_3439_cast_fp16 = slice_by_index(begin = var_3439_begin_0, end = var_3439_end_0, end_mask = var_3439_end_mask_0, x = new_x0_30_cast_fp16)[name = string("op_3439_cast_fp16")];
            string x_299_pad_type_0 = const()[name = string("x_299_pad_type_0"), val = string("valid")];
            int32 x_299_groups_0 = const()[name = string("x_299_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_299_strides_0 = const()[name = string("x_299_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_299_pad_0 = const()[name = string("x_299_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_299_dilations_0 = const()[name = string("x_299_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359436672))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359445952))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_299_cast_fp16 = conv(dilations = x_299_dilations_0, groups = x_299_groups_0, pad = x_299_pad_0, pad_type = x_299_pad_type_0, strides = x_299_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_30_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<int32, [3]> input1_92_perm_0 = const()[name = string("input1_92_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_301_axes_0 = const()[name = string("x_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359446528)))];
            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359448640)))];
            tensor<fp16, [1, 4, 1024]> input1_92_cast_fp16 = transpose(perm = input1_92_perm_0, x = x_299_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 4, 1024]> x_301_cast_fp16 = layer_norm(axes = x_301_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input1_92_cast_fp16)[name = string("x_301_cast_fp16")];
            tensor<int32, [3]> input2_60_perm_0 = const()[name = string("input2_60_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_60_cast_fp16 = transpose(perm = input2_60_perm_0, x = x_301_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [1, 1024, 4]> var_3454_cast_fp16 = silu(x = input2_60_cast_fp16)[name = string("op_3454_cast_fp16")];
            string x_303_pad_type_0 = const()[name = string("x_303_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_303_strides_0 = const()[name = string("x_303_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_303_pad_0 = const()[name = string("x_303_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_303_dilations_0 = const()[name = string("x_303_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_303_groups_0 = const()[name = string("x_303_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359450752))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360499392))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_303_cast_fp16 = conv(dilations = x_303_dilations_0, groups = x_303_groups_0, pad = x_303_pad_0, pad_type = x_303_pad_type_0, strides = x_303_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_3454_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [3]> input3_32_perm_0 = const()[name = string("input3_32_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_32_cast_fp16 = transpose(perm = input3_32_perm_0, x = x_303_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [1, 4, 1024]> input1_94_cast_fp16 = add(x = input0_181_cast_fp16, y = input3_32_cast_fp16)[name = string("input1_94_cast_fp16")];
            tensor<int32, [1]> input0_187_axes_0 = const()[name = string("input0_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360499968)))];
            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360502080)))];
            tensor<fp16, [1, 4, 1024]> input0_187_cast_fp16 = layer_norm(axes = input0_187_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input1_94_cast_fp16)[name = string("input0_187_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360504192))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364698560))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_187_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3475_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("op_3475_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364699136))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368893504))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = var_3475_cast_fp16)[name = string("linear_135_cast_fp16")];
            fp16 var_3480_to_fp16 = const()[name = string("op_3480_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3481_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3480_to_fp16)[name = string("op_3481_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_62_cast_fp16 = add(x = input1_94_cast_fp16, y = var_3481_cast_fp16)[name = string("input2_62_cast_fp16")];
            tensor<int32, [1]> input0_189_axes_0 = const()[name = string("input0_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368894080)))];
            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368896192)))];
            tensor<fp16, [1, 4, 1024]> input0_189_cast_fp16 = layer_norm(axes = input0_189_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input2_62_cast_fp16)[name = string("input0_189_cast_fp16")];
            tensor<int32, [4]> cache29_1_begin_0 = const()[name = string("cache29_1_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> cache29_1_end_0 = const()[name = string("cache29_1_end_0"), val = tensor<int32, [4]>([16, 1, 56, 1024])];
            tensor<bool, [4]> cache29_1_end_mask_0 = const()[name = string("cache29_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache29_1_squeeze_mask_0 = const()[name = string("cache29_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache29_1_cast_fp16 = slice_by_index(begin = cache29_1_begin_0, end = cache29_1_end_0, end_mask = cache29_1_end_mask_0, squeeze_mask = cache29_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache29_1_cast_fp16")];
            tensor<int32, [4]> cache30_1_begin_0 = const()[name = string("cache30_1_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> cache30_1_end_0 = const()[name = string("cache30_1_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
            tensor<bool, [4]> cache30_1_end_mask_0 = const()[name = string("cache30_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache30_1_squeeze_mask_0 = const()[name = string("cache30_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache30_1_cast_fp16 = slice_by_index(begin = cache30_1_begin_0, end = cache30_1_end_0, end_mask = cache30_1_end_mask_0, squeeze_mask = cache30_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache30_1_cast_fp16")];
            tensor<int32, [1]> input_189_axes_0 = const()[name = string("input_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368898304)))];
            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368900416)))];
            tensor<fp16, [1, 4, 1024]> input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input0_189_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368902528))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373096896))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_189_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3510_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373097472))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377291840))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = var_3510_cast_fp16)[name = string("linear_137_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3516_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_193_cast_fp16 = add(x = input0_189_cast_fp16, y = var_3516_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<int32, [1]> key_32_axes_0 = const()[name = string("key_32_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377292416)))];
            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377294528)))];
            tensor<fp16, [1, 4, 1024]> key_32_cast_fp16 = layer_norm(axes = key_32_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_32_cast_fp16")];
            bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_195_cast_fp16 = concat(axis = var_64, interleave = input_195_interleave_0, values = (cache29_1_cast_fp16, key_32_cast_fp16))[name = string("input_195_cast_fp16")];
            tensor<int32, [3]> var_3538_begin_0 = const()[name = string("op_3538_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3538_end_0 = const()[name = string("op_3538_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_3538_end_mask_0 = const()[name = string("op_3538_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_3538_cast_fp16 = slice_by_index(begin = var_3538_begin_0, end = var_3538_end_0, end_mask = var_3538_end_mask_0, x = cache29_1_cast_fp16)[name = string("op_3538_cast_fp16")];
            bool var_3544_interleave_0 = const()[name = string("op_3544_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_3544_cast_fp16 = concat(axis = var_64, interleave = var_3544_interleave_0, values = (var_3538_cast_fp16, key_32_cast_fp16))[name = string("op_3544_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377296640))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378345280))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_32_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_3548 = const()[name = string("op_3548"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_32_cast_fp16 = reshape(shape = var_3548, x = linear_138_cast_fp16)[name = string("q_32_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378345856))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379394496))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<int32, [4]> var_3552 = const()[name = string("op_3552"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_32_cast_fp16 = reshape(shape = var_3552, x = linear_139_cast_fp16)[name = string("k_32_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379395072))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380443712))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3556 = const()[name = string("op_3556"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_32_cast_fp16 = reshape(shape = var_3556, x = linear_140_cast_fp16)[name = string("v_32_cast_fp16")];
            tensor<int32, [4]> value_34_perm_0 = const()[name = string("value_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380444288)))];
            tensor<fp16, [1, 4, 8, 128]> var_3568_cast_fp16 = add(x = q_32_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3568_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380446400)))];
            tensor<fp16, [1, 4, 8, 128]> var_3570_cast_fp16 = add(x = q_32_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3570_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_32_perm_0 = const()[name = string("q_with_bias_v_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_311_transpose_x_0 = const()[name = string("x_311_transpose_x_0"), val = bool(false)];
            bool x_311_transpose_y_0 = const()[name = string("x_311_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_3572_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380448512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380570432))))[name = string("op_3572_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_32_cast_fp16 = transpose(perm = q_with_bias_v_32_perm_0, x = var_3570_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [1, 8, 4, 119]> x_311_cast_fp16 = matmul(transpose_x = x_311_transpose_x_0, transpose_y = x_311_transpose_y_0, x = q_with_bias_v_32_cast_fp16, y = op_3572_to_fp16_palettized)[name = string("x_311_cast_fp16")];
            tensor<int32, [8]> x0_34_pad_0 = const()[name = string("x0_34_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_34_mode_0 = const()[name = string("x0_34_mode_0"), val = string("constant")];
            fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_34_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x0_34_mode_0, pad = x0_34_pad_0, x = x_311_cast_fp16)[name = string("x0_34_cast_fp16")];
            tensor<int32, [4]> var_3580 = const()[name = string("op_3580"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_32_cast_fp16 = reshape(shape = var_3580, x = x0_34_cast_fp16)[name = string("x1_32_cast_fp16")];
            tensor<int32, [4]> var_3584_begin_0 = const()[name = string("op_3584_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3584_end_0 = const()[name = string("op_3584_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_3584_end_mask_0 = const()[name = string("op_3584_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_3584_cast_fp16 = slice_by_index(begin = var_3584_begin_0, end = var_3584_end_0, end_mask = var_3584_end_mask_0, x = x1_32_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<int32, [4]> var_3585 = const()[name = string("op_3585"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_32_cast_fp16 = reshape(shape = var_3585, x = var_3584_cast_fp16)[name = string("matrix_bd_32_cast_fp16")];
            bool matrix_ac_32_transpose_x_0 = const()[name = string("matrix_ac_32_transpose_x_0"), val = bool(false)];
            bool matrix_ac_32_transpose_y_0 = const()[name = string("matrix_ac_32_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_32_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 8, 4, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3568_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_32_cast_fp16 = matmul(transpose_x = matrix_ac_32_transpose_x_0, transpose_y = matrix_ac_32_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_32_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_32_begin_0 = const()[name = string("matrix_bd0_32_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_32_end_0 = const()[name = string("matrix_bd0_32_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_32_end_mask_0 = const()[name = string("matrix_bd0_32_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_32_cast_fp16 = slice_by_index(begin = matrix_bd0_32_begin_0, end = matrix_bd0_32_end_0, end_mask = matrix_bd0_32_end_mask_0, x = matrix_bd_32_cast_fp16)[name = string("matrix_bd0_32_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3594_cast_fp16 = add(x = matrix_ac_32_cast_fp16, y = matrix_bd0_32_cast_fp16)[name = string("op_3594_cast_fp16")];
            fp16 _inversed_scores_32_y_0_to_fp16 = const()[name = string("_inversed_scores_32_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_32_cast_fp16 = mul(x = var_3594_cast_fp16, y = _inversed_scores_32_y_0_to_fp16)[name = string("_inversed_scores_32_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_32_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_32_cast_fp16, cond = mask0_4)[name = string("scores0_32_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3600_cast_fp16 = softmax(axis = var_56, x = scores0_32_cast_fp16)[name = string("op_3600_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_191_cast_fp16 = select(a = var_30_to_fp16, b = var_3600_cast_fp16, cond = mask0_4)[name = string("input0_191_cast_fp16")];
            bool x2_32_transpose_x_0 = const()[name = string("x2_32_transpose_x_0"), val = bool(false)];
            bool x2_32_transpose_y_0 = const()[name = string("x2_32_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_34_cast_fp16 = transpose(perm = value_34_perm_0, x = v_32_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 8, 4, 128]> x2_32_cast_fp16 = matmul(transpose_x = x2_32_transpose_x_0, transpose_y = x2_32_transpose_y_0, x = input0_191_cast_fp16, y = value_34_cast_fp16)[name = string("x2_32_cast_fp16")];
            tensor<int32, [4]> var_3604_perm_0 = const()[name = string("op_3604_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3605 = const()[name = string("op_3605"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_3604_cast_fp16 = transpose(perm = var_3604_perm_0, x = x2_32_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 4, 1024]> input1_96_cast_fp16 = reshape(shape = var_3605, x = var_3604_cast_fp16)[name = string("input1_96_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380571008))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381619648))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input1_96_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_193_cast_fp16 = add(x = input_193_cast_fp16, y = linear_142_cast_fp16)[name = string("input0_193_cast_fp16")];
            tensor<int32, [1]> x_315_axes_0 = const()[name = string("x_315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381620224)))];
            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381622336)))];
            tensor<fp16, [1, 4, 1024]> x_315_cast_fp16 = layer_norm(axes = x_315_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input0_193_cast_fp16)[name = string("x_315_cast_fp16")];
            tensor<int32, [3]> input_197_perm_0 = const()[name = string("input_197_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_195_pad_type_0 = const()[name = string("input0_195_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_195_strides_0 = const()[name = string("input0_195_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_195_pad_0 = const()[name = string("input0_195_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_195_dilations_0 = const()[name = string("input0_195_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_195_groups_0 = const()[name = string("input0_195_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381624448))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383721664))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_197_cast_fp16 = transpose(perm = input_197_perm_0, x = x_315_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 2048, 4]> input0_195_cast_fp16 = conv(dilations = input0_195_dilations_0, groups = input0_195_groups_0, pad = input0_195_pad_0, pad_type = input0_195_pad_type_0, strides = input0_195_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = string("input0_195_cast_fp16")];
            int32 x_317_split_num_splits_0 = const()[name = string("x_317_split_num_splits_0"), val = int32(2)];
            int32 x_317_split_axis_0 = const()[name = string("x_317_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_317_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_317_split_cast_fp16_1 = split(axis = x_317_split_axis_0, num_splits = x_317_split_num_splits_0, x = input0_195_cast_fp16)[name = string("x_317_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_317_split_1_sigmoid_cast_fp16 = sigmoid(x = x_317_split_cast_fp16_1)[name = string("x_317_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_317_cast_fp16 = mul(x = x_317_split_cast_fp16_0, y = x_317_split_1_sigmoid_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_197_cast_fp16 = select(a = var_30_to_fp16, b = x_317_cast_fp16, cond = var_570)[name = string("input0_197_cast_fp16")];
            bool new_x0_32_interleave_0 = const()[name = string("new_x0_32_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_32_cast_fp16 = concat(axis = var_56, interleave = new_x0_32_interleave_0, values = (cache30_1_cast_fp16, input0_197_cast_fp16))[name = string("new_x0_32_cast_fp16")];
            tensor<int32, [3]> var_3643_begin_0 = const()[name = string("op_3643_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3643_end_0 = const()[name = string("op_3643_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_3643_end_mask_0 = const()[name = string("op_3643_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_3643_cast_fp16 = slice_by_index(begin = var_3643_begin_0, end = var_3643_end_0, end_mask = var_3643_end_mask_0, x = new_x0_32_cast_fp16)[name = string("op_3643_cast_fp16")];
            string x_319_pad_type_0 = const()[name = string("x_319_pad_type_0"), val = string("valid")];
            int32 x_319_groups_0 = const()[name = string("x_319_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_319_strides_0 = const()[name = string("x_319_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_319_pad_0 = const()[name = string("x_319_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_319_dilations_0 = const()[name = string("x_319_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383722240))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383731520))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_319_cast_fp16 = conv(dilations = x_319_dilations_0, groups = x_319_groups_0, pad = x_319_pad_0, pad_type = x_319_pad_type_0, strides = x_319_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_32_cast_fp16)[name = string("x_319_cast_fp16")];
            tensor<int32, [3]> input1_98_perm_0 = const()[name = string("input1_98_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_321_axes_0 = const()[name = string("x_321_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383732096)))];
            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383734208)))];
            tensor<fp16, [1, 4, 1024]> input1_98_cast_fp16 = transpose(perm = input1_98_perm_0, x = x_319_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 4, 1024]> x_321_cast_fp16 = layer_norm(axes = x_321_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input1_98_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [3]> input2_64_perm_0 = const()[name = string("input2_64_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_64_cast_fp16 = transpose(perm = input2_64_perm_0, x = x_321_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 1024, 4]> var_3658_cast_fp16 = silu(x = input2_64_cast_fp16)[name = string("op_3658_cast_fp16")];
            string x_323_pad_type_0 = const()[name = string("x_323_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_323_strides_0 = const()[name = string("x_323_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_323_pad_0 = const()[name = string("x_323_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_323_dilations_0 = const()[name = string("x_323_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_323_groups_0 = const()[name = string("x_323_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383736320))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384784960))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_323_cast_fp16 = conv(dilations = x_323_dilations_0, groups = x_323_groups_0, pad = x_323_pad_0, pad_type = x_323_pad_type_0, strides = x_323_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_3658_cast_fp16)[name = string("x_323_cast_fp16")];
            tensor<int32, [3]> input3_34_perm_0 = const()[name = string("input3_34_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_34_cast_fp16 = transpose(perm = input3_34_perm_0, x = x_323_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [1, 4, 1024]> input1_100_cast_fp16 = add(x = input0_193_cast_fp16, y = input3_34_cast_fp16)[name = string("input1_100_cast_fp16")];
            tensor<int32, [1]> input0_199_axes_0 = const()[name = string("input0_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384785536)))];
            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384787648)))];
            tensor<fp16, [1, 4, 1024]> input0_199_cast_fp16 = layer_norm(axes = input0_199_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input1_100_cast_fp16)[name = string("input0_199_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384789760))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388984128))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_199_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3679_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388984704))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393179072))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = var_3679_cast_fp16)[name = string("linear_144_cast_fp16")];
            fp16 var_3684_to_fp16 = const()[name = string("op_3684_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3685_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3684_to_fp16)[name = string("op_3685_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_66_cast_fp16 = add(x = input1_100_cast_fp16, y = var_3685_cast_fp16)[name = string("input2_66_cast_fp16")];
            tensor<int32, [1]> input0_201_axes_0 = const()[name = string("input0_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393179648)))];
            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393181760)))];
            tensor<fp16, [1, 4, 1024]> input0_201_cast_fp16 = layer_norm(axes = input0_201_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input2_66_cast_fp16)[name = string("input0_201_cast_fp16")];
            tensor<int32, [4]> cache31_1_begin_0 = const()[name = string("cache31_1_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> cache31_1_end_0 = const()[name = string("cache31_1_end_0"), val = tensor<int32, [4]>([17, 1, 56, 1024])];
            tensor<bool, [4]> cache31_1_end_mask_0 = const()[name = string("cache31_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache31_1_squeeze_mask_0 = const()[name = string("cache31_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache31_1_cast_fp16 = slice_by_index(begin = cache31_1_begin_0, end = cache31_1_end_0, end_mask = cache31_1_end_mask_0, squeeze_mask = cache31_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache31_1_cast_fp16")];
            tensor<int32, [4]> cache32_1_begin_0 = const()[name = string("cache32_1_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> cache32_1_end_0 = const()[name = string("cache32_1_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
            tensor<bool, [4]> cache32_1_end_mask_0 = const()[name = string("cache32_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache32_1_squeeze_mask_0 = const()[name = string("cache32_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache32_1_cast_fp16 = slice_by_index(begin = cache32_1_begin_0, end = cache32_1_end_0, end_mask = cache32_1_end_mask_0, squeeze_mask = cache32_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache32_1_cast_fp16")];
            tensor<int32, [1]> input_201_axes_0 = const()[name = string("input_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393183872)))];
            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393185984)))];
            tensor<fp16, [1, 4, 1024]> input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input0_201_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393188096))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397382464))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_201_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3714_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("op_3714_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397383040))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401577408))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = var_3714_cast_fp16)[name = string("linear_146_cast_fp16")];
            fp16 var_3719_to_fp16 = const()[name = string("op_3719_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3720_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3719_to_fp16)[name = string("op_3720_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_205_cast_fp16 = add(x = input0_201_cast_fp16, y = var_3720_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<int32, [1]> key_34_axes_0 = const()[name = string("key_34_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401577984)))];
            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401580096)))];
            tensor<fp16, [1, 4, 1024]> key_34_cast_fp16 = layer_norm(axes = key_34_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_205_cast_fp16)[name = string("key_34_cast_fp16")];
            bool input_207_interleave_0 = const()[name = string("input_207_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_207_cast_fp16 = concat(axis = var_64, interleave = input_207_interleave_0, values = (cache31_1_cast_fp16, key_34_cast_fp16))[name = string("input_207_cast_fp16")];
            tensor<int32, [3]> var_3742_begin_0 = const()[name = string("op_3742_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3742_end_0 = const()[name = string("op_3742_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_3742_end_mask_0 = const()[name = string("op_3742_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_3742_cast_fp16 = slice_by_index(begin = var_3742_begin_0, end = var_3742_end_0, end_mask = var_3742_end_mask_0, x = cache31_1_cast_fp16)[name = string("op_3742_cast_fp16")];
            bool var_3748_interleave_0 = const()[name = string("op_3748_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_3748_cast_fp16 = concat(axis = var_64, interleave = var_3748_interleave_0, values = (var_3742_cast_fp16, key_34_cast_fp16))[name = string("op_3748_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401582208))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402630848))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_34_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3752 = const()[name = string("op_3752"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_34_cast_fp16 = reshape(shape = var_3752, x = linear_147_cast_fp16)[name = string("q_34_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402631424))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403680064))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3756 = const()[name = string("op_3756"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_34_cast_fp16 = reshape(shape = var_3756, x = linear_148_cast_fp16)[name = string("k_34_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403680640))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404729280))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3760 = const()[name = string("op_3760"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_34_cast_fp16 = reshape(shape = var_3760, x = linear_149_cast_fp16)[name = string("v_34_cast_fp16")];
            tensor<int32, [4]> value_36_perm_0 = const()[name = string("value_36_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404729856)))];
            tensor<fp16, [1, 4, 8, 128]> var_3772_cast_fp16 = add(x = q_34_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3772_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404731968)))];
            tensor<fp16, [1, 4, 8, 128]> var_3774_cast_fp16 = add(x = q_34_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3774_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_34_perm_0 = const()[name = string("q_with_bias_v_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_331_transpose_x_0 = const()[name = string("x_331_transpose_x_0"), val = bool(false)];
            bool x_331_transpose_y_0 = const()[name = string("x_331_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_3776_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404734080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404856000))))[name = string("op_3776_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_34_cast_fp16 = transpose(perm = q_with_bias_v_34_perm_0, x = var_3774_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [1, 8, 4, 119]> x_331_cast_fp16 = matmul(transpose_x = x_331_transpose_x_0, transpose_y = x_331_transpose_y_0, x = q_with_bias_v_34_cast_fp16, y = op_3776_to_fp16_palettized)[name = string("x_331_cast_fp16")];
            tensor<int32, [8]> x0_36_pad_0 = const()[name = string("x0_36_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_36_mode_0 = const()[name = string("x0_36_mode_0"), val = string("constant")];
            fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_36_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x0_36_mode_0, pad = x0_36_pad_0, x = x_331_cast_fp16)[name = string("x0_36_cast_fp16")];
            tensor<int32, [4]> var_3784 = const()[name = string("op_3784"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_34_cast_fp16 = reshape(shape = var_3784, x = x0_36_cast_fp16)[name = string("x1_34_cast_fp16")];
            tensor<int32, [4]> var_3788_begin_0 = const()[name = string("op_3788_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3788_end_0 = const()[name = string("op_3788_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_3788_end_mask_0 = const()[name = string("op_3788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_3788_cast_fp16 = slice_by_index(begin = var_3788_begin_0, end = var_3788_end_0, end_mask = var_3788_end_mask_0, x = x1_34_cast_fp16)[name = string("op_3788_cast_fp16")];
            tensor<int32, [4]> var_3789 = const()[name = string("op_3789"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_34_cast_fp16 = reshape(shape = var_3789, x = var_3788_cast_fp16)[name = string("matrix_bd_34_cast_fp16")];
            bool matrix_ac_34_transpose_x_0 = const()[name = string("matrix_ac_34_transpose_x_0"), val = bool(false)];
            bool matrix_ac_34_transpose_y_0 = const()[name = string("matrix_ac_34_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_34_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [1, 8, 4, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3772_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_34_cast_fp16 = matmul(transpose_x = matrix_ac_34_transpose_x_0, transpose_y = matrix_ac_34_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_34_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_34_begin_0 = const()[name = string("matrix_bd0_34_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_34_end_0 = const()[name = string("matrix_bd0_34_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_34_end_mask_0 = const()[name = string("matrix_bd0_34_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_34_cast_fp16 = slice_by_index(begin = matrix_bd0_34_begin_0, end = matrix_bd0_34_end_0, end_mask = matrix_bd0_34_end_mask_0, x = matrix_bd_34_cast_fp16)[name = string("matrix_bd0_34_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3798_cast_fp16 = add(x = matrix_ac_34_cast_fp16, y = matrix_bd0_34_cast_fp16)[name = string("op_3798_cast_fp16")];
            fp16 _inversed_scores_34_y_0_to_fp16 = const()[name = string("_inversed_scores_34_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_34_cast_fp16 = mul(x = var_3798_cast_fp16, y = _inversed_scores_34_y_0_to_fp16)[name = string("_inversed_scores_34_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_34_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_34_cast_fp16, cond = mask0_4)[name = string("scores0_34_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_3804_cast_fp16 = softmax(axis = var_56, x = scores0_34_cast_fp16)[name = string("op_3804_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_203_cast_fp16 = select(a = var_30_to_fp16, b = var_3804_cast_fp16, cond = mask0_4)[name = string("input0_203_cast_fp16")];
            bool x2_34_transpose_x_0 = const()[name = string("x2_34_transpose_x_0"), val = bool(false)];
            bool x2_34_transpose_y_0 = const()[name = string("x2_34_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_36_cast_fp16 = transpose(perm = value_36_perm_0, x = v_34_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 8, 4, 128]> x2_34_cast_fp16 = matmul(transpose_x = x2_34_transpose_x_0, transpose_y = x2_34_transpose_y_0, x = input0_203_cast_fp16, y = value_36_cast_fp16)[name = string("x2_34_cast_fp16")];
            tensor<int32, [4]> var_3808_perm_0 = const()[name = string("op_3808_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3809 = const()[name = string("op_3809"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_3808_cast_fp16 = transpose(perm = var_3808_perm_0, x = x2_34_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 1024]> input1_102_cast_fp16 = reshape(shape = var_3809, x = var_3808_cast_fp16)[name = string("input1_102_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404856576))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405905216))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input1_102_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_205_cast_fp16 = add(x = input_205_cast_fp16, y = linear_151_cast_fp16)[name = string("input0_205_cast_fp16")];
            tensor<int32, [1]> x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405905792)))];
            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405907904)))];
            tensor<fp16, [1, 4, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input0_205_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<int32, [3]> input_209_perm_0 = const()[name = string("input_209_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_207_pad_type_0 = const()[name = string("input0_207_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_207_strides_0 = const()[name = string("input0_207_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_207_pad_0 = const()[name = string("input0_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_207_dilations_0 = const()[name = string("input0_207_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_207_groups_0 = const()[name = string("input0_207_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405910016))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408007232))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_209_cast_fp16 = transpose(perm = input_209_perm_0, x = x_335_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 2048, 4]> input0_207_cast_fp16 = conv(dilations = input0_207_dilations_0, groups = input0_207_groups_0, pad = input0_207_pad_0, pad_type = input0_207_pad_type_0, strides = input0_207_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_209_cast_fp16)[name = string("input0_207_cast_fp16")];
            int32 x_337_split_num_splits_0 = const()[name = string("x_337_split_num_splits_0"), val = int32(2)];
            int32 x_337_split_axis_0 = const()[name = string("x_337_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_337_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_337_split_cast_fp16_1 = split(axis = x_337_split_axis_0, num_splits = x_337_split_num_splits_0, x = input0_207_cast_fp16)[name = string("x_337_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_337_split_1_sigmoid_cast_fp16 = sigmoid(x = x_337_split_cast_fp16_1)[name = string("x_337_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_337_cast_fp16 = mul(x = x_337_split_cast_fp16_0, y = x_337_split_1_sigmoid_cast_fp16)[name = string("x_337_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_209_cast_fp16 = select(a = var_30_to_fp16, b = x_337_cast_fp16, cond = var_570)[name = string("input0_209_cast_fp16")];
            bool new_x0_34_interleave_0 = const()[name = string("new_x0_34_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_34_cast_fp16 = concat(axis = var_56, interleave = new_x0_34_interleave_0, values = (cache32_1_cast_fp16, input0_209_cast_fp16))[name = string("new_x0_34_cast_fp16")];
            tensor<int32, [3]> var_3847_begin_0 = const()[name = string("op_3847_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3847_end_0 = const()[name = string("op_3847_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_3847_end_mask_0 = const()[name = string("op_3847_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_3847_cast_fp16 = slice_by_index(begin = var_3847_begin_0, end = var_3847_end_0, end_mask = var_3847_end_mask_0, x = new_x0_34_cast_fp16)[name = string("op_3847_cast_fp16")];
            string x_339_pad_type_0 = const()[name = string("x_339_pad_type_0"), val = string("valid")];
            int32 x_339_groups_0 = const()[name = string("x_339_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_339_strides_0 = const()[name = string("x_339_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_339_pad_0 = const()[name = string("x_339_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_339_dilations_0 = const()[name = string("x_339_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408007808))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408017088))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_339_cast_fp16 = conv(dilations = x_339_dilations_0, groups = x_339_groups_0, pad = x_339_pad_0, pad_type = x_339_pad_type_0, strides = x_339_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_34_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [3]> input1_104_perm_0 = const()[name = string("input1_104_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_341_axes_0 = const()[name = string("x_341_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408017664)))];
            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408019776)))];
            tensor<fp16, [1, 4, 1024]> input1_104_cast_fp16 = transpose(perm = input1_104_perm_0, x = x_339_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 1024]> x_341_cast_fp16 = layer_norm(axes = x_341_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input1_104_cast_fp16)[name = string("x_341_cast_fp16")];
            tensor<int32, [3]> input2_68_perm_0 = const()[name = string("input2_68_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_68_cast_fp16 = transpose(perm = input2_68_perm_0, x = x_341_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 1024, 4]> var_3862_cast_fp16 = silu(x = input2_68_cast_fp16)[name = string("op_3862_cast_fp16")];
            string x_343_pad_type_0 = const()[name = string("x_343_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_343_strides_0 = const()[name = string("x_343_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_343_pad_0 = const()[name = string("x_343_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_343_dilations_0 = const()[name = string("x_343_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_343_groups_0 = const()[name = string("x_343_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408021888))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409070528))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_343_cast_fp16 = conv(dilations = x_343_dilations_0, groups = x_343_groups_0, pad = x_343_pad_0, pad_type = x_343_pad_type_0, strides = x_343_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_3862_cast_fp16)[name = string("x_343_cast_fp16")];
            tensor<int32, [3]> input3_36_perm_0 = const()[name = string("input3_36_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_36_cast_fp16 = transpose(perm = input3_36_perm_0, x = x_343_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 4, 1024]> input1_106_cast_fp16 = add(x = input0_205_cast_fp16, y = input3_36_cast_fp16)[name = string("input1_106_cast_fp16")];
            tensor<int32, [1]> input0_211_axes_0 = const()[name = string("input0_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409071104)))];
            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409073216)))];
            tensor<fp16, [1, 4, 1024]> input0_211_cast_fp16 = layer_norm(axes = input0_211_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input1_106_cast_fp16)[name = string("input0_211_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409075328))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413269696))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_211_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3883_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("op_3883_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413270272))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417464640))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = var_3883_cast_fp16)[name = string("linear_153_cast_fp16")];
            fp16 var_3888_to_fp16 = const()[name = string("op_3888_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3889_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_3888_to_fp16)[name = string("op_3889_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_70_cast_fp16 = add(x = input1_106_cast_fp16, y = var_3889_cast_fp16)[name = string("input2_70_cast_fp16")];
            tensor<int32, [1]> input0_213_axes_0 = const()[name = string("input0_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417465216)))];
            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417467328)))];
            tensor<fp16, [1, 4, 1024]> input0_213_cast_fp16 = layer_norm(axes = input0_213_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input2_70_cast_fp16)[name = string("input0_213_cast_fp16")];
            tensor<int32, [4]> cache33_1_begin_0 = const()[name = string("cache33_1_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> cache33_1_end_0 = const()[name = string("cache33_1_end_0"), val = tensor<int32, [4]>([18, 1, 56, 1024])];
            tensor<bool, [4]> cache33_1_end_mask_0 = const()[name = string("cache33_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache33_1_squeeze_mask_0 = const()[name = string("cache33_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache33_1_cast_fp16 = slice_by_index(begin = cache33_1_begin_0, end = cache33_1_end_0, end_mask = cache33_1_end_mask_0, squeeze_mask = cache33_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache33_1_cast_fp16")];
            tensor<int32, [4]> cache34_1_begin_0 = const()[name = string("cache34_1_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> cache34_1_end_0 = const()[name = string("cache34_1_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
            tensor<bool, [4]> cache34_1_end_mask_0 = const()[name = string("cache34_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache34_1_squeeze_mask_0 = const()[name = string("cache34_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache34_1_cast_fp16 = slice_by_index(begin = cache34_1_begin_0, end = cache34_1_end_0, end_mask = cache34_1_end_mask_0, squeeze_mask = cache34_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache34_1_cast_fp16")];
            tensor<int32, [1]> input_213_axes_0 = const()[name = string("input_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417469440)))];
            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417471552)))];
            tensor<fp16, [1, 4, 1024]> input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input0_213_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417473664))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421668032))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_154_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_3918_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("op_3918_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421668608))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425862976))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = var_3918_cast_fp16)[name = string("linear_155_cast_fp16")];
            fp16 var_3923_to_fp16 = const()[name = string("op_3923_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_3924_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_3923_to_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_217_cast_fp16 = add(x = input0_213_cast_fp16, y = var_3924_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<int32, [1]> key_36_axes_0 = const()[name = string("key_36_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425863552)))];
            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425865664)))];
            tensor<fp16, [1, 4, 1024]> key_36_cast_fp16 = layer_norm(axes = key_36_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_217_cast_fp16)[name = string("key_36_cast_fp16")];
            bool input_219_interleave_0 = const()[name = string("input_219_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_219_cast_fp16 = concat(axis = var_64, interleave = input_219_interleave_0, values = (cache33_1_cast_fp16, key_36_cast_fp16))[name = string("input_219_cast_fp16")];
            tensor<int32, [3]> var_3946_begin_0 = const()[name = string("op_3946_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_3946_end_0 = const()[name = string("op_3946_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_3946_end_mask_0 = const()[name = string("op_3946_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_3946_cast_fp16 = slice_by_index(begin = var_3946_begin_0, end = var_3946_end_0, end_mask = var_3946_end_mask_0, x = cache33_1_cast_fp16)[name = string("op_3946_cast_fp16")];
            bool var_3952_interleave_0 = const()[name = string("op_3952_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_3952_cast_fp16 = concat(axis = var_64, interleave = var_3952_interleave_0, values = (var_3946_cast_fp16, key_36_cast_fp16))[name = string("op_3952_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425867776))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426916416))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_36_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3956 = const()[name = string("op_3956"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_36_cast_fp16 = reshape(shape = var_3956, x = linear_156_cast_fp16)[name = string("q_36_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426916992))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427965632))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<int32, [4]> var_3960 = const()[name = string("op_3960"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_36_cast_fp16 = reshape(shape = var_3960, x = linear_157_cast_fp16)[name = string("k_36_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427966208))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429014848))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_158_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<int32, [4]> var_3964 = const()[name = string("op_3964"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_36_cast_fp16 = reshape(shape = var_3964, x = linear_158_cast_fp16)[name = string("v_36_cast_fp16")];
            tensor<int32, [4]> value_38_perm_0 = const()[name = string("value_38_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429015424)))];
            tensor<fp16, [1, 4, 8, 128]> var_3976_cast_fp16 = add(x = q_36_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_3976_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429017536)))];
            tensor<fp16, [1, 4, 8, 128]> var_3978_cast_fp16 = add(x = q_36_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_3978_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_36_perm_0 = const()[name = string("q_with_bias_v_36_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_3980_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429019648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429141568))))[name = string("op_3980_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_36_cast_fp16 = transpose(perm = q_with_bias_v_36_perm_0, x = var_3978_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [1, 8, 4, 119]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = q_with_bias_v_36_cast_fp16, y = op_3980_to_fp16_palettized)[name = string("x_351_cast_fp16")];
            tensor<int32, [8]> x0_38_pad_0 = const()[name = string("x0_38_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_38_mode_0 = const()[name = string("x0_38_mode_0"), val = string("constant")];
            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_38_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x0_38_mode_0, pad = x0_38_pad_0, x = x_351_cast_fp16)[name = string("x0_38_cast_fp16")];
            tensor<int32, [4]> var_3988 = const()[name = string("op_3988"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_36_cast_fp16 = reshape(shape = var_3988, x = x0_38_cast_fp16)[name = string("x1_36_cast_fp16")];
            tensor<int32, [4]> var_3992_begin_0 = const()[name = string("op_3992_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3992_end_0 = const()[name = string("op_3992_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_3992_end_mask_0 = const()[name = string("op_3992_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_3992_cast_fp16 = slice_by_index(begin = var_3992_begin_0, end = var_3992_end_0, end_mask = var_3992_end_mask_0, x = x1_36_cast_fp16)[name = string("op_3992_cast_fp16")];
            tensor<int32, [4]> var_3993 = const()[name = string("op_3993"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_36_cast_fp16 = reshape(shape = var_3993, x = var_3992_cast_fp16)[name = string("matrix_bd_36_cast_fp16")];
            bool matrix_ac_36_transpose_x_0 = const()[name = string("matrix_ac_36_transpose_x_0"), val = bool(false)];
            bool matrix_ac_36_transpose_y_0 = const()[name = string("matrix_ac_36_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_36_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [1, 8, 4, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_3976_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_36_cast_fp16 = matmul(transpose_x = matrix_ac_36_transpose_x_0, transpose_y = matrix_ac_36_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_36_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_36_begin_0 = const()[name = string("matrix_bd0_36_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_36_end_0 = const()[name = string("matrix_bd0_36_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_36_end_mask_0 = const()[name = string("matrix_bd0_36_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_36_cast_fp16 = slice_by_index(begin = matrix_bd0_36_begin_0, end = matrix_bd0_36_end_0, end_mask = matrix_bd0_36_end_mask_0, x = matrix_bd_36_cast_fp16)[name = string("matrix_bd0_36_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4002_cast_fp16 = add(x = matrix_ac_36_cast_fp16, y = matrix_bd0_36_cast_fp16)[name = string("op_4002_cast_fp16")];
            fp16 _inversed_scores_36_y_0_to_fp16 = const()[name = string("_inversed_scores_36_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_36_cast_fp16 = mul(x = var_4002_cast_fp16, y = _inversed_scores_36_y_0_to_fp16)[name = string("_inversed_scores_36_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_36_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_36_cast_fp16, cond = mask0_4)[name = string("scores0_36_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4008_cast_fp16 = softmax(axis = var_56, x = scores0_36_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_215_cast_fp16 = select(a = var_30_to_fp16, b = var_4008_cast_fp16, cond = mask0_4)[name = string("input0_215_cast_fp16")];
            bool x2_36_transpose_x_0 = const()[name = string("x2_36_transpose_x_0"), val = bool(false)];
            bool x2_36_transpose_y_0 = const()[name = string("x2_36_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_38_cast_fp16 = transpose(perm = value_38_perm_0, x = v_36_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [1, 8, 4, 128]> x2_36_cast_fp16 = matmul(transpose_x = x2_36_transpose_x_0, transpose_y = x2_36_transpose_y_0, x = input0_215_cast_fp16, y = value_38_cast_fp16)[name = string("x2_36_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4013 = const()[name = string("op_4013"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = x2_36_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 1024]> input1_108_cast_fp16 = reshape(shape = var_4013, x = var_4012_cast_fp16)[name = string("input1_108_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429142144))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430190784))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input1_108_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_217_cast_fp16 = add(x = input_217_cast_fp16, y = linear_160_cast_fp16)[name = string("input0_217_cast_fp16")];
            tensor<int32, [1]> x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430191360)))];
            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430193472)))];
            tensor<fp16, [1, 4, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input0_217_cast_fp16)[name = string("x_355_cast_fp16")];
            tensor<int32, [3]> input_221_perm_0 = const()[name = string("input_221_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_219_pad_type_0 = const()[name = string("input0_219_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_219_strides_0 = const()[name = string("input0_219_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_219_pad_0 = const()[name = string("input0_219_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_219_dilations_0 = const()[name = string("input0_219_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_219_groups_0 = const()[name = string("input0_219_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430195584))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432292800))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_221_cast_fp16 = transpose(perm = input_221_perm_0, x = x_355_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 2048, 4]> input0_219_cast_fp16 = conv(dilations = input0_219_dilations_0, groups = input0_219_groups_0, pad = input0_219_pad_0, pad_type = input0_219_pad_type_0, strides = input0_219_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("input0_219_cast_fp16")];
            int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)];
            int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input0_219_cast_fp16)[name = string("x_357_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_221_cast_fp16 = select(a = var_30_to_fp16, b = x_357_cast_fp16, cond = var_570)[name = string("input0_221_cast_fp16")];
            bool new_x0_36_interleave_0 = const()[name = string("new_x0_36_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_36_cast_fp16 = concat(axis = var_56, interleave = new_x0_36_interleave_0, values = (cache34_1_cast_fp16, input0_221_cast_fp16))[name = string("new_x0_36_cast_fp16")];
            tensor<int32, [3]> var_4051_begin_0 = const()[name = string("op_4051_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4051_end_0 = const()[name = string("op_4051_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_4051_end_mask_0 = const()[name = string("op_4051_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_4051_cast_fp16 = slice_by_index(begin = var_4051_begin_0, end = var_4051_end_0, end_mask = var_4051_end_mask_0, x = new_x0_36_cast_fp16)[name = string("op_4051_cast_fp16")];
            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_17_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432293376))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432302656))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_36_cast_fp16)[name = string("x_359_cast_fp16")];
            tensor<int32, [3]> input1_110_perm_0 = const()[name = string("input1_110_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432303232)))];
            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432305344)))];
            tensor<fp16, [1, 4, 1024]> input1_110_cast_fp16 = transpose(perm = input1_110_perm_0, x = x_359_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 4, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input1_110_cast_fp16)[name = string("x_361_cast_fp16")];
            tensor<int32, [3]> input2_72_perm_0 = const()[name = string("input2_72_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_72_cast_fp16 = transpose(perm = input2_72_perm_0, x = x_361_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 1024, 4]> var_4066_cast_fp16 = silu(x = input2_72_cast_fp16)[name = string("op_4066_cast_fp16")];
            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432307456))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433356096))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_4066_cast_fp16)[name = string("x_363_cast_fp16")];
            tensor<int32, [3]> input3_38_perm_0 = const()[name = string("input3_38_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_38_cast_fp16 = transpose(perm = input3_38_perm_0, x = x_363_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 4, 1024]> input1_112_cast_fp16 = add(x = input0_217_cast_fp16, y = input3_38_cast_fp16)[name = string("input1_112_cast_fp16")];
            tensor<int32, [1]> input0_223_axes_0 = const()[name = string("input0_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433356672)))];
            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433358784)))];
            tensor<fp16, [1, 4, 1024]> input0_223_cast_fp16 = layer_norm(axes = input0_223_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input1_112_cast_fp16)[name = string("input0_223_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433360896))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437555264))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_161_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_223_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4087_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("op_4087_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437555840))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441750208))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = var_4087_cast_fp16)[name = string("linear_162_cast_fp16")];
            fp16 var_4092_to_fp16 = const()[name = string("op_4092_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4093_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4092_to_fp16)[name = string("op_4093_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_74_cast_fp16 = add(x = input1_112_cast_fp16, y = var_4093_cast_fp16)[name = string("input2_74_cast_fp16")];
            tensor<int32, [1]> input0_225_axes_0 = const()[name = string("input0_225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441750784)))];
            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441752896)))];
            tensor<fp16, [1, 4, 1024]> input0_225_cast_fp16 = layer_norm(axes = input0_225_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input2_74_cast_fp16)[name = string("input0_225_cast_fp16")];
            tensor<int32, [4]> cache35_1_begin_0 = const()[name = string("cache35_1_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> cache35_1_end_0 = const()[name = string("cache35_1_end_0"), val = tensor<int32, [4]>([19, 1, 56, 1024])];
            tensor<bool, [4]> cache35_1_end_mask_0 = const()[name = string("cache35_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache35_1_squeeze_mask_0 = const()[name = string("cache35_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache35_1_cast_fp16 = slice_by_index(begin = cache35_1_begin_0, end = cache35_1_end_0, end_mask = cache35_1_end_mask_0, squeeze_mask = cache35_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache35_1_cast_fp16")];
            tensor<int32, [4]> cache36_1_begin_0 = const()[name = string("cache36_1_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> cache36_1_end_0 = const()[name = string("cache36_1_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
            tensor<bool, [4]> cache36_1_end_mask_0 = const()[name = string("cache36_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache36_1_squeeze_mask_0 = const()[name = string("cache36_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache36_1_cast_fp16 = slice_by_index(begin = cache36_1_begin_0, end = cache36_1_end_0, end_mask = cache36_1_end_mask_0, squeeze_mask = cache36_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache36_1_cast_fp16")];
            tensor<int32, [1]> input_225_axes_0 = const()[name = string("input_225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441755008)))];
            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441757120)))];
            tensor<fp16, [1, 4, 1024]> input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input0_225_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441759232))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445953600))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4122_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445954176))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450148544))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = var_4122_cast_fp16)[name = string("linear_164_cast_fp16")];
            fp16 var_4127_to_fp16 = const()[name = string("op_4127_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4128_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4127_to_fp16)[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_229_cast_fp16 = add(x = input0_225_cast_fp16, y = var_4128_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<int32, [1]> key_38_axes_0 = const()[name = string("key_38_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450149120)))];
            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450151232)))];
            tensor<fp16, [1, 4, 1024]> key_38_cast_fp16 = layer_norm(axes = key_38_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_229_cast_fp16)[name = string("key_38_cast_fp16")];
            bool input_231_interleave_0 = const()[name = string("input_231_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_231_cast_fp16 = concat(axis = var_64, interleave = input_231_interleave_0, values = (cache35_1_cast_fp16, key_38_cast_fp16))[name = string("input_231_cast_fp16")];
            tensor<int32, [3]> var_4150_begin_0 = const()[name = string("op_4150_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_4150_end_0 = const()[name = string("op_4150_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_4150_end_mask_0 = const()[name = string("op_4150_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_4150_cast_fp16 = slice_by_index(begin = var_4150_begin_0, end = var_4150_end_0, end_mask = var_4150_end_mask_0, x = cache35_1_cast_fp16)[name = string("op_4150_cast_fp16")];
            bool var_4156_interleave_0 = const()[name = string("op_4156_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_4156_cast_fp16 = concat(axis = var_64, interleave = var_4156_interleave_0, values = (var_4150_cast_fp16, key_38_cast_fp16))[name = string("op_4156_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450153344))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451201984))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_38_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_4160 = const()[name = string("op_4160"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_38_cast_fp16 = reshape(shape = var_4160, x = linear_165_cast_fp16)[name = string("q_38_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451202560))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452251200))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<int32, [4]> var_4164 = const()[name = string("op_4164"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_38_cast_fp16 = reshape(shape = var_4164, x = linear_166_cast_fp16)[name = string("k_38_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452251776))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453300416))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_167_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<int32, [4]> var_4168 = const()[name = string("op_4168"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_38_cast_fp16 = reshape(shape = var_4168, x = linear_167_cast_fp16)[name = string("v_38_cast_fp16")];
            tensor<int32, [4]> value_40_perm_0 = const()[name = string("value_40_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453300992)))];
            tensor<fp16, [1, 4, 8, 128]> var_4180_cast_fp16 = add(x = q_38_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4180_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453303104)))];
            tensor<fp16, [1, 4, 8, 128]> var_4182_cast_fp16 = add(x = q_38_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4182_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_38_perm_0 = const()[name = string("q_with_bias_v_38_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_4184_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453305216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453427136))))[name = string("op_4184_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_38_cast_fp16 = transpose(perm = q_with_bias_v_38_perm_0, x = var_4182_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 8, 4, 119]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_38_cast_fp16, y = op_4184_to_fp16_palettized)[name = string("x_371_cast_fp16")];
            tensor<int32, [8]> x0_40_pad_0 = const()[name = string("x0_40_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_40_mode_0 = const()[name = string("x0_40_mode_0"), val = string("constant")];
            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_40_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x0_40_mode_0, pad = x0_40_pad_0, x = x_371_cast_fp16)[name = string("x0_40_cast_fp16")];
            tensor<int32, [4]> var_4192 = const()[name = string("op_4192"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_38_cast_fp16 = reshape(shape = var_4192, x = x0_40_cast_fp16)[name = string("x1_38_cast_fp16")];
            tensor<int32, [4]> var_4196_begin_0 = const()[name = string("op_4196_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4196_end_0 = const()[name = string("op_4196_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_4196_end_mask_0 = const()[name = string("op_4196_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_4196_cast_fp16 = slice_by_index(begin = var_4196_begin_0, end = var_4196_end_0, end_mask = var_4196_end_mask_0, x = x1_38_cast_fp16)[name = string("op_4196_cast_fp16")];
            tensor<int32, [4]> var_4197 = const()[name = string("op_4197"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_38_cast_fp16 = reshape(shape = var_4197, x = var_4196_cast_fp16)[name = string("matrix_bd_38_cast_fp16")];
            bool matrix_ac_38_transpose_x_0 = const()[name = string("matrix_ac_38_transpose_x_0"), val = bool(false)];
            bool matrix_ac_38_transpose_y_0 = const()[name = string("matrix_ac_38_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_38_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [1, 8, 4, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4180_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_38_cast_fp16 = matmul(transpose_x = matrix_ac_38_transpose_x_0, transpose_y = matrix_ac_38_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_38_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_38_begin_0 = const()[name = string("matrix_bd0_38_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_38_end_0 = const()[name = string("matrix_bd0_38_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_38_end_mask_0 = const()[name = string("matrix_bd0_38_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_38_cast_fp16 = slice_by_index(begin = matrix_bd0_38_begin_0, end = matrix_bd0_38_end_0, end_mask = matrix_bd0_38_end_mask_0, x = matrix_bd_38_cast_fp16)[name = string("matrix_bd0_38_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4206_cast_fp16 = add(x = matrix_ac_38_cast_fp16, y = matrix_bd0_38_cast_fp16)[name = string("op_4206_cast_fp16")];
            fp16 _inversed_scores_38_y_0_to_fp16 = const()[name = string("_inversed_scores_38_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_38_cast_fp16 = mul(x = var_4206_cast_fp16, y = _inversed_scores_38_y_0_to_fp16)[name = string("_inversed_scores_38_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_38_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_38_cast_fp16, cond = mask0_4)[name = string("scores0_38_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4212_cast_fp16 = softmax(axis = var_56, x = scores0_38_cast_fp16)[name = string("op_4212_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_227_cast_fp16 = select(a = var_30_to_fp16, b = var_4212_cast_fp16, cond = mask0_4)[name = string("input0_227_cast_fp16")];
            bool x2_38_transpose_x_0 = const()[name = string("x2_38_transpose_x_0"), val = bool(false)];
            bool x2_38_transpose_y_0 = const()[name = string("x2_38_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_40_cast_fp16 = transpose(perm = value_40_perm_0, x = v_38_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [1, 8, 4, 128]> x2_38_cast_fp16 = matmul(transpose_x = x2_38_transpose_x_0, transpose_y = x2_38_transpose_y_0, x = input0_227_cast_fp16, y = value_40_cast_fp16)[name = string("x2_38_cast_fp16")];
            tensor<int32, [4]> var_4216_perm_0 = const()[name = string("op_4216_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4217 = const()[name = string("op_4217"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_4216_cast_fp16 = transpose(perm = var_4216_perm_0, x = x2_38_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [1, 4, 1024]> input1_114_cast_fp16 = reshape(shape = var_4217, x = var_4216_cast_fp16)[name = string("input1_114_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453427712))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454476352))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_palettized, x = input1_114_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_229_cast_fp16 = add(x = input_229_cast_fp16, y = linear_169_cast_fp16)[name = string("input0_229_cast_fp16")];
            tensor<int32, [1]> x_375_axes_0 = const()[name = string("x_375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454476928)))];
            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454479040)))];
            tensor<fp16, [1, 4, 1024]> x_375_cast_fp16 = layer_norm(axes = x_375_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input0_229_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [3]> input_233_perm_0 = const()[name = string("input_233_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_231_pad_type_0 = const()[name = string("input0_231_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_231_strides_0 = const()[name = string("input0_231_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_231_pad_0 = const()[name = string("input0_231_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_231_dilations_0 = const()[name = string("input0_231_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_231_groups_0 = const()[name = string("input0_231_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454481152))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456578368))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_233_cast_fp16 = transpose(perm = input_233_perm_0, x = x_375_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 2048, 4]> input0_231_cast_fp16 = conv(dilations = input0_231_dilations_0, groups = input0_231_groups_0, pad = input0_231_pad_0, pad_type = input0_231_pad_type_0, strides = input0_231_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_233_cast_fp16)[name = string("input0_231_cast_fp16")];
            int32 x_377_split_num_splits_0 = const()[name = string("x_377_split_num_splits_0"), val = int32(2)];
            int32 x_377_split_axis_0 = const()[name = string("x_377_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_377_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_377_split_cast_fp16_1 = split(axis = x_377_split_axis_0, num_splits = x_377_split_num_splits_0, x = input0_231_cast_fp16)[name = string("x_377_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_377_split_1_sigmoid_cast_fp16 = sigmoid(x = x_377_split_cast_fp16_1)[name = string("x_377_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_377_cast_fp16 = mul(x = x_377_split_cast_fp16_0, y = x_377_split_1_sigmoid_cast_fp16)[name = string("x_377_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_233_cast_fp16 = select(a = var_30_to_fp16, b = x_377_cast_fp16, cond = var_570)[name = string("input0_233_cast_fp16")];
            bool new_x0_38_interleave_0 = const()[name = string("new_x0_38_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_38_cast_fp16 = concat(axis = var_56, interleave = new_x0_38_interleave_0, values = (cache36_1_cast_fp16, input0_233_cast_fp16))[name = string("new_x0_38_cast_fp16")];
            tensor<int32, [3]> var_4255_begin_0 = const()[name = string("op_4255_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4255_end_0 = const()[name = string("op_4255_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_4255_end_mask_0 = const()[name = string("op_4255_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, x = new_x0_38_cast_fp16)[name = string("op_4255_cast_fp16")];
            string x_379_pad_type_0 = const()[name = string("x_379_pad_type_0"), val = string("valid")];
            int32 x_379_groups_0 = const()[name = string("x_379_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_379_strides_0 = const()[name = string("x_379_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_379_pad_0 = const()[name = string("x_379_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_379_dilations_0 = const()[name = string("x_379_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_18_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456578944))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456588224))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_379_cast_fp16 = conv(dilations = x_379_dilations_0, groups = x_379_groups_0, pad = x_379_pad_0, pad_type = x_379_pad_type_0, strides = x_379_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_38_cast_fp16)[name = string("x_379_cast_fp16")];
            tensor<int32, [3]> input1_116_perm_0 = const()[name = string("input1_116_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456588800)))];
            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456590912)))];
            tensor<fp16, [1, 4, 1024]> input1_116_cast_fp16 = transpose(perm = input1_116_perm_0, x = x_379_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input1_116_cast_fp16)[name = string("x_381_cast_fp16")];
            tensor<int32, [3]> input2_76_perm_0 = const()[name = string("input2_76_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_76_cast_fp16 = transpose(perm = input2_76_perm_0, x = x_381_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 1024, 4]> var_4270_cast_fp16 = silu(x = input2_76_cast_fp16)[name = string("op_4270_cast_fp16")];
            string x_383_pad_type_0 = const()[name = string("x_383_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_383_strides_0 = const()[name = string("x_383_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_383_pad_0 = const()[name = string("x_383_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_383_dilations_0 = const()[name = string("x_383_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_383_groups_0 = const()[name = string("x_383_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456593024))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457641664))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_383_cast_fp16 = conv(dilations = x_383_dilations_0, groups = x_383_groups_0, pad = x_383_pad_0, pad_type = x_383_pad_type_0, strides = x_383_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_4270_cast_fp16)[name = string("x_383_cast_fp16")];
            tensor<int32, [3]> input3_40_perm_0 = const()[name = string("input3_40_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_40_cast_fp16 = transpose(perm = input3_40_perm_0, x = x_383_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 1024]> input1_118_cast_fp16 = add(x = input0_229_cast_fp16, y = input3_40_cast_fp16)[name = string("input1_118_cast_fp16")];
            tensor<int32, [1]> input0_235_axes_0 = const()[name = string("input0_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457642240)))];
            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457644352)))];
            tensor<fp16, [1, 4, 1024]> input0_235_cast_fp16 = layer_norm(axes = input0_235_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input1_118_cast_fp16)[name = string("input0_235_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457646464))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461840832))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_235_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4291_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("op_4291_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461841408))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466035776))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_palettized, x = var_4291_cast_fp16)[name = string("linear_171_cast_fp16")];
            fp16 var_4296_to_fp16 = const()[name = string("op_4296_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4297_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4296_to_fp16)[name = string("op_4297_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_78_cast_fp16 = add(x = input1_118_cast_fp16, y = var_4297_cast_fp16)[name = string("input2_78_cast_fp16")];
            tensor<int32, [1]> input0_237_axes_0 = const()[name = string("input0_237_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466036352)))];
            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466038464)))];
            tensor<fp16, [1, 4, 1024]> input0_237_cast_fp16 = layer_norm(axes = input0_237_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input2_78_cast_fp16)[name = string("input0_237_cast_fp16")];
            tensor<int32, [4]> cache37_1_begin_0 = const()[name = string("cache37_1_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> cache37_1_end_0 = const()[name = string("cache37_1_end_0"), val = tensor<int32, [4]>([20, 1, 56, 1024])];
            tensor<bool, [4]> cache37_1_end_mask_0 = const()[name = string("cache37_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache37_1_squeeze_mask_0 = const()[name = string("cache37_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache37_1_cast_fp16 = slice_by_index(begin = cache37_1_begin_0, end = cache37_1_end_0, end_mask = cache37_1_end_mask_0, squeeze_mask = cache37_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache37_1_cast_fp16")];
            tensor<int32, [4]> cache38_1_begin_0 = const()[name = string("cache38_1_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> cache38_1_end_0 = const()[name = string("cache38_1_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
            tensor<bool, [4]> cache38_1_end_mask_0 = const()[name = string("cache38_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache38_1_squeeze_mask_0 = const()[name = string("cache38_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache38_1_cast_fp16 = slice_by_index(begin = cache38_1_begin_0, end = cache38_1_end_0, end_mask = cache38_1_end_mask_0, squeeze_mask = cache38_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache38_1_cast_fp16")];
            tensor<int32, [1]> input_237_axes_0 = const()[name = string("input_237_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466040576)))];
            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466042688)))];
            tensor<fp16, [1, 4, 1024]> input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input0_237_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466044800))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470239168))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4326_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_4326_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470239744))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474434112))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_palettized, x = var_4326_cast_fp16)[name = string("linear_173_cast_fp16")];
            fp16 var_4331_to_fp16 = const()[name = string("op_4331_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4332_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4331_to_fp16)[name = string("op_4332_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_241_cast_fp16 = add(x = input0_237_cast_fp16, y = var_4332_cast_fp16)[name = string("input_241_cast_fp16")];
            tensor<int32, [1]> key_40_axes_0 = const()[name = string("key_40_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474434688)))];
            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474436800)))];
            tensor<fp16, [1, 4, 1024]> key_40_cast_fp16 = layer_norm(axes = key_40_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_241_cast_fp16)[name = string("key_40_cast_fp16")];
            bool input_243_interleave_0 = const()[name = string("input_243_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_243_cast_fp16 = concat(axis = var_64, interleave = input_243_interleave_0, values = (cache37_1_cast_fp16, key_40_cast_fp16))[name = string("input_243_cast_fp16")];
            tensor<int32, [3]> var_4354_begin_0 = const()[name = string("op_4354_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_4354_end_0 = const()[name = string("op_4354_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_4354_end_mask_0 = const()[name = string("op_4354_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_4354_cast_fp16 = slice_by_index(begin = var_4354_begin_0, end = var_4354_end_0, end_mask = var_4354_end_mask_0, x = cache37_1_cast_fp16)[name = string("op_4354_cast_fp16")];
            bool var_4360_interleave_0 = const()[name = string("op_4360_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_4360_cast_fp16 = concat(axis = var_64, interleave = var_4360_interleave_0, values = (var_4354_cast_fp16, key_40_cast_fp16))[name = string("op_4360_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474438912))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475487552))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_palettized, x = key_40_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_40_cast_fp16 = reshape(shape = var_4364, x = linear_174_cast_fp16)[name = string("q_40_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475488128))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476536768))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<int32, [4]> var_4368 = const()[name = string("op_4368"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_40_cast_fp16 = reshape(shape = var_4368, x = linear_175_cast_fp16)[name = string("k_40_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476537344))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477585984))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_176_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [4]> var_4372 = const()[name = string("op_4372"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_40_cast_fp16 = reshape(shape = var_4372, x = linear_176_cast_fp16)[name = string("v_40_cast_fp16")];
            tensor<int32, [4]> value_42_perm_0 = const()[name = string("value_42_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477586560)))];
            tensor<fp16, [1, 4, 8, 128]> var_4384_cast_fp16 = add(x = q_40_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4384_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477588672)))];
            tensor<fp16, [1, 4, 8, 128]> var_4386_cast_fp16 = add(x = q_40_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4386_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_40_perm_0 = const()[name = string("q_with_bias_v_40_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_391_transpose_x_0 = const()[name = string("x_391_transpose_x_0"), val = bool(false)];
            bool x_391_transpose_y_0 = const()[name = string("x_391_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_4388_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477590784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477712704))))[name = string("op_4388_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_40_cast_fp16 = transpose(perm = q_with_bias_v_40_perm_0, x = var_4386_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 8, 4, 119]> x_391_cast_fp16 = matmul(transpose_x = x_391_transpose_x_0, transpose_y = x_391_transpose_y_0, x = q_with_bias_v_40_cast_fp16, y = op_4388_to_fp16_palettized)[name = string("x_391_cast_fp16")];
            tensor<int32, [8]> x0_42_pad_0 = const()[name = string("x0_42_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_42_mode_0 = const()[name = string("x0_42_mode_0"), val = string("constant")];
            fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_42_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x0_42_mode_0, pad = x0_42_pad_0, x = x_391_cast_fp16)[name = string("x0_42_cast_fp16")];
            tensor<int32, [4]> var_4396 = const()[name = string("op_4396"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_40_cast_fp16 = reshape(shape = var_4396, x = x0_42_cast_fp16)[name = string("x1_40_cast_fp16")];
            tensor<int32, [4]> var_4400_begin_0 = const()[name = string("op_4400_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4400_end_0 = const()[name = string("op_4400_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_4400_end_mask_0 = const()[name = string("op_4400_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, x = x1_40_cast_fp16)[name = string("op_4400_cast_fp16")];
            tensor<int32, [4]> var_4401 = const()[name = string("op_4401"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_40_cast_fp16 = reshape(shape = var_4401, x = var_4400_cast_fp16)[name = string("matrix_bd_40_cast_fp16")];
            bool matrix_ac_40_transpose_x_0 = const()[name = string("matrix_ac_40_transpose_x_0"), val = bool(false)];
            bool matrix_ac_40_transpose_y_0 = const()[name = string("matrix_ac_40_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_40_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [1, 8, 4, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4384_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_40_cast_fp16 = matmul(transpose_x = matrix_ac_40_transpose_x_0, transpose_y = matrix_ac_40_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_40_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_40_begin_0 = const()[name = string("matrix_bd0_40_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_40_end_0 = const()[name = string("matrix_bd0_40_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_40_end_mask_0 = const()[name = string("matrix_bd0_40_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_40_cast_fp16 = slice_by_index(begin = matrix_bd0_40_begin_0, end = matrix_bd0_40_end_0, end_mask = matrix_bd0_40_end_mask_0, x = matrix_bd_40_cast_fp16)[name = string("matrix_bd0_40_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4410_cast_fp16 = add(x = matrix_ac_40_cast_fp16, y = matrix_bd0_40_cast_fp16)[name = string("op_4410_cast_fp16")];
            fp16 _inversed_scores_40_y_0_to_fp16 = const()[name = string("_inversed_scores_40_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_40_cast_fp16 = mul(x = var_4410_cast_fp16, y = _inversed_scores_40_y_0_to_fp16)[name = string("_inversed_scores_40_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_40_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_40_cast_fp16, cond = mask0_4)[name = string("scores0_40_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4416_cast_fp16 = softmax(axis = var_56, x = scores0_40_cast_fp16)[name = string("op_4416_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_239_cast_fp16 = select(a = var_30_to_fp16, b = var_4416_cast_fp16, cond = mask0_4)[name = string("input0_239_cast_fp16")];
            bool x2_40_transpose_x_0 = const()[name = string("x2_40_transpose_x_0"), val = bool(false)];
            bool x2_40_transpose_y_0 = const()[name = string("x2_40_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_42_cast_fp16 = transpose(perm = value_42_perm_0, x = v_40_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [1, 8, 4, 128]> x2_40_cast_fp16 = matmul(transpose_x = x2_40_transpose_x_0, transpose_y = x2_40_transpose_y_0, x = input0_239_cast_fp16, y = value_42_cast_fp16)[name = string("x2_40_cast_fp16")];
            tensor<int32, [4]> var_4420_perm_0 = const()[name = string("op_4420_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4421 = const()[name = string("op_4421"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_4420_cast_fp16 = transpose(perm = var_4420_perm_0, x = x2_40_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [1, 4, 1024]> input1_120_cast_fp16 = reshape(shape = var_4421, x = var_4420_cast_fp16)[name = string("input1_120_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477713280))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478761920))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_palettized, x = input1_120_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_241_cast_fp16 = add(x = input_241_cast_fp16, y = linear_178_cast_fp16)[name = string("input0_241_cast_fp16")];
            tensor<int32, [1]> x_395_axes_0 = const()[name = string("x_395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478762496)))];
            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478764608)))];
            tensor<fp16, [1, 4, 1024]> x_395_cast_fp16 = layer_norm(axes = x_395_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input0_241_cast_fp16)[name = string("x_395_cast_fp16")];
            tensor<int32, [3]> input_245_perm_0 = const()[name = string("input_245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_243_pad_type_0 = const()[name = string("input0_243_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_243_strides_0 = const()[name = string("input0_243_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_243_pad_0 = const()[name = string("input0_243_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_243_dilations_0 = const()[name = string("input0_243_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_243_groups_0 = const()[name = string("input0_243_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478766720))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480863936))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_245_cast_fp16 = transpose(perm = input_245_perm_0, x = x_395_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [1, 2048, 4]> input0_243_cast_fp16 = conv(dilations = input0_243_dilations_0, groups = input0_243_groups_0, pad = input0_243_pad_0, pad_type = input0_243_pad_type_0, strides = input0_243_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_245_cast_fp16)[name = string("input0_243_cast_fp16")];
            int32 x_397_split_num_splits_0 = const()[name = string("x_397_split_num_splits_0"), val = int32(2)];
            int32 x_397_split_axis_0 = const()[name = string("x_397_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_397_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_397_split_cast_fp16_1 = split(axis = x_397_split_axis_0, num_splits = x_397_split_num_splits_0, x = input0_243_cast_fp16)[name = string("x_397_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_397_split_1_sigmoid_cast_fp16 = sigmoid(x = x_397_split_cast_fp16_1)[name = string("x_397_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_397_cast_fp16 = mul(x = x_397_split_cast_fp16_0, y = x_397_split_1_sigmoid_cast_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_245_cast_fp16 = select(a = var_30_to_fp16, b = x_397_cast_fp16, cond = var_570)[name = string("input0_245_cast_fp16")];
            bool new_x0_40_interleave_0 = const()[name = string("new_x0_40_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_40_cast_fp16 = concat(axis = var_56, interleave = new_x0_40_interleave_0, values = (cache38_1_cast_fp16, input0_245_cast_fp16))[name = string("new_x0_40_cast_fp16")];
            tensor<int32, [3]> var_4459_begin_0 = const()[name = string("op_4459_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4459_end_0 = const()[name = string("op_4459_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_4459_end_mask_0 = const()[name = string("op_4459_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_4459_cast_fp16 = slice_by_index(begin = var_4459_begin_0, end = var_4459_end_0, end_mask = var_4459_end_mask_0, x = new_x0_40_cast_fp16)[name = string("op_4459_cast_fp16")];
            string x_399_pad_type_0 = const()[name = string("x_399_pad_type_0"), val = string("valid")];
            int32 x_399_groups_0 = const()[name = string("x_399_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_399_strides_0 = const()[name = string("x_399_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_399_dilations_0 = const()[name = string("x_399_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_19_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480864512))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480873792))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_399_cast_fp16 = conv(dilations = x_399_dilations_0, groups = x_399_groups_0, pad = x_399_pad_0, pad_type = x_399_pad_type_0, strides = x_399_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_40_cast_fp16)[name = string("x_399_cast_fp16")];
            tensor<int32, [3]> input1_122_perm_0 = const()[name = string("input1_122_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_401_axes_0 = const()[name = string("x_401_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480874368)))];
            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480876480)))];
            tensor<fp16, [1, 4, 1024]> input1_122_cast_fp16 = transpose(perm = input1_122_perm_0, x = x_399_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 1024]> x_401_cast_fp16 = layer_norm(axes = x_401_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input1_122_cast_fp16)[name = string("x_401_cast_fp16")];
            tensor<int32, [3]> input2_80_perm_0 = const()[name = string("input2_80_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_80_cast_fp16 = transpose(perm = input2_80_perm_0, x = x_401_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 1024, 4]> var_4474_cast_fp16 = silu(x = input2_80_cast_fp16)[name = string("op_4474_cast_fp16")];
            string x_403_pad_type_0 = const()[name = string("x_403_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_403_strides_0 = const()[name = string("x_403_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_403_pad_0 = const()[name = string("x_403_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_403_dilations_0 = const()[name = string("x_403_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_403_groups_0 = const()[name = string("x_403_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480878592))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481927232))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_403_cast_fp16 = conv(dilations = x_403_dilations_0, groups = x_403_groups_0, pad = x_403_pad_0, pad_type = x_403_pad_type_0, strides = x_403_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_4474_cast_fp16)[name = string("x_403_cast_fp16")];
            tensor<int32, [3]> input3_42_perm_0 = const()[name = string("input3_42_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_42_cast_fp16 = transpose(perm = input3_42_perm_0, x = x_403_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 4, 1024]> input1_124_cast_fp16 = add(x = input0_241_cast_fp16, y = input3_42_cast_fp16)[name = string("input1_124_cast_fp16")];
            tensor<int32, [1]> input0_247_axes_0 = const()[name = string("input0_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481927808)))];
            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481929920)))];
            tensor<fp16, [1, 4, 1024]> input0_247_cast_fp16 = layer_norm(axes = input0_247_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input1_124_cast_fp16)[name = string("input0_247_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481932032))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486126400))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_179_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_247_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4495_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_4495_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486126976))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490321344))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_palettized, x = var_4495_cast_fp16)[name = string("linear_180_cast_fp16")];
            fp16 var_4500_to_fp16 = const()[name = string("op_4500_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4501_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4500_to_fp16)[name = string("op_4501_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_82_cast_fp16 = add(x = input1_124_cast_fp16, y = var_4501_cast_fp16)[name = string("input2_82_cast_fp16")];
            tensor<int32, [1]> input0_249_axes_0 = const()[name = string("input0_249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490321920)))];
            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490324032)))];
            tensor<fp16, [1, 4, 1024]> input0_249_cast_fp16 = layer_norm(axes = input0_249_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input2_82_cast_fp16)[name = string("input0_249_cast_fp16")];
            tensor<int32, [4]> cache39_1_begin_0 = const()[name = string("cache39_1_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> cache39_1_end_0 = const()[name = string("cache39_1_end_0"), val = tensor<int32, [4]>([21, 1, 56, 1024])];
            tensor<bool, [4]> cache39_1_end_mask_0 = const()[name = string("cache39_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache39_1_squeeze_mask_0 = const()[name = string("cache39_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache39_1_cast_fp16 = slice_by_index(begin = cache39_1_begin_0, end = cache39_1_end_0, end_mask = cache39_1_end_mask_0, squeeze_mask = cache39_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache39_1_cast_fp16")];
            tensor<int32, [4]> cache40_1_begin_0 = const()[name = string("cache40_1_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> cache40_1_end_0 = const()[name = string("cache40_1_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
            tensor<bool, [4]> cache40_1_end_mask_0 = const()[name = string("cache40_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache40_1_squeeze_mask_0 = const()[name = string("cache40_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache40_1_cast_fp16 = slice_by_index(begin = cache40_1_begin_0, end = cache40_1_end_0, end_mask = cache40_1_end_mask_0, squeeze_mask = cache40_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache40_1_cast_fp16")];
            tensor<int32, [1]> input_249_axes_0 = const()[name = string("input_249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326144)))];
            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490328256)))];
            tensor<fp16, [1, 4, 1024]> input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input0_249_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490330368))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494524736))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_palettized, x = input_249_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4530_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("op_4530_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494525312))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498719680))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_palettized, x = var_4530_cast_fp16)[name = string("linear_182_cast_fp16")];
            fp16 var_4535_to_fp16 = const()[name = string("op_4535_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4536_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4535_to_fp16)[name = string("op_4536_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_253_cast_fp16 = add(x = input0_249_cast_fp16, y = var_4536_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<int32, [1]> key_42_axes_0 = const()[name = string("key_42_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498720256)))];
            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498722368)))];
            tensor<fp16, [1, 4, 1024]> key_42_cast_fp16 = layer_norm(axes = key_42_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_253_cast_fp16)[name = string("key_42_cast_fp16")];
            bool input_255_interleave_0 = const()[name = string("input_255_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_255_cast_fp16 = concat(axis = var_64, interleave = input_255_interleave_0, values = (cache39_1_cast_fp16, key_42_cast_fp16))[name = string("input_255_cast_fp16")];
            tensor<int32, [3]> var_4558_begin_0 = const()[name = string("op_4558_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_4558_end_0 = const()[name = string("op_4558_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_4558_end_mask_0 = const()[name = string("op_4558_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_4558_cast_fp16 = slice_by_index(begin = var_4558_begin_0, end = var_4558_end_0, end_mask = var_4558_end_mask_0, x = cache39_1_cast_fp16)[name = string("op_4558_cast_fp16")];
            bool var_4564_interleave_0 = const()[name = string("op_4564_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_4564_cast_fp16 = concat(axis = var_64, interleave = var_4564_interleave_0, values = (var_4558_cast_fp16, key_42_cast_fp16))[name = string("op_4564_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498724480))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499773120))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_palettized, x = key_42_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_4568 = const()[name = string("op_4568"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_42_cast_fp16 = reshape(shape = var_4568, x = linear_183_cast_fp16)[name = string("q_42_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499773696))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500822336))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_palettized, x = input_255_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> var_4572 = const()[name = string("op_4572"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_42_cast_fp16 = reshape(shape = var_4572, x = linear_184_cast_fp16)[name = string("k_42_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500822912))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501871552))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_185_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_palettized, x = input_255_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<int32, [4]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_42_cast_fp16 = reshape(shape = var_4576, x = linear_185_cast_fp16)[name = string("v_42_cast_fp16")];
            tensor<int32, [4]> value_44_perm_0 = const()[name = string("value_44_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501872128)))];
            tensor<fp16, [1, 4, 8, 128]> var_4588_cast_fp16 = add(x = q_42_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501874240)))];
            tensor<fp16, [1, 4, 8, 128]> var_4590_cast_fp16 = add(x = q_42_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4590_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_42_perm_0 = const()[name = string("q_with_bias_v_42_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_411_transpose_x_0 = const()[name = string("x_411_transpose_x_0"), val = bool(false)];
            bool x_411_transpose_y_0 = const()[name = string("x_411_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_4592_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501876352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501998272))))[name = string("op_4592_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_42_cast_fp16 = transpose(perm = q_with_bias_v_42_perm_0, x = var_4590_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 8, 4, 119]> x_411_cast_fp16 = matmul(transpose_x = x_411_transpose_x_0, transpose_y = x_411_transpose_y_0, x = q_with_bias_v_42_cast_fp16, y = op_4592_to_fp16_palettized)[name = string("x_411_cast_fp16")];
            tensor<int32, [8]> x0_44_pad_0 = const()[name = string("x0_44_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_44_mode_0 = const()[name = string("x0_44_mode_0"), val = string("constant")];
            fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_44_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x0_44_mode_0, pad = x0_44_pad_0, x = x_411_cast_fp16)[name = string("x0_44_cast_fp16")];
            tensor<int32, [4]> var_4600 = const()[name = string("op_4600"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_42_cast_fp16 = reshape(shape = var_4600, x = x0_44_cast_fp16)[name = string("x1_42_cast_fp16")];
            tensor<int32, [4]> var_4604_begin_0 = const()[name = string("op_4604_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4604_end_0 = const()[name = string("op_4604_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_4604_end_mask_0 = const()[name = string("op_4604_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_4604_cast_fp16 = slice_by_index(begin = var_4604_begin_0, end = var_4604_end_0, end_mask = var_4604_end_mask_0, x = x1_42_cast_fp16)[name = string("op_4604_cast_fp16")];
            tensor<int32, [4]> var_4605 = const()[name = string("op_4605"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_42_cast_fp16 = reshape(shape = var_4605, x = var_4604_cast_fp16)[name = string("matrix_bd_42_cast_fp16")];
            bool matrix_ac_42_transpose_x_0 = const()[name = string("matrix_ac_42_transpose_x_0"), val = bool(false)];
            bool matrix_ac_42_transpose_y_0 = const()[name = string("matrix_ac_42_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_42_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 8, 4, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4588_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_42_cast_fp16 = matmul(transpose_x = matrix_ac_42_transpose_x_0, transpose_y = matrix_ac_42_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_42_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_42_begin_0 = const()[name = string("matrix_bd0_42_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_42_end_0 = const()[name = string("matrix_bd0_42_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_42_end_mask_0 = const()[name = string("matrix_bd0_42_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_42_cast_fp16 = slice_by_index(begin = matrix_bd0_42_begin_0, end = matrix_bd0_42_end_0, end_mask = matrix_bd0_42_end_mask_0, x = matrix_bd_42_cast_fp16)[name = string("matrix_bd0_42_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4614_cast_fp16 = add(x = matrix_ac_42_cast_fp16, y = matrix_bd0_42_cast_fp16)[name = string("op_4614_cast_fp16")];
            fp16 _inversed_scores_42_y_0_to_fp16 = const()[name = string("_inversed_scores_42_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_42_cast_fp16 = mul(x = var_4614_cast_fp16, y = _inversed_scores_42_y_0_to_fp16)[name = string("_inversed_scores_42_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_42_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_42_cast_fp16, cond = mask0_4)[name = string("scores0_42_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4620_cast_fp16 = softmax(axis = var_56, x = scores0_42_cast_fp16)[name = string("op_4620_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_251_cast_fp16 = select(a = var_30_to_fp16, b = var_4620_cast_fp16, cond = mask0_4)[name = string("input0_251_cast_fp16")];
            bool x2_42_transpose_x_0 = const()[name = string("x2_42_transpose_x_0"), val = bool(false)];
            bool x2_42_transpose_y_0 = const()[name = string("x2_42_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_44_cast_fp16 = transpose(perm = value_44_perm_0, x = v_42_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [1, 8, 4, 128]> x2_42_cast_fp16 = matmul(transpose_x = x2_42_transpose_x_0, transpose_y = x2_42_transpose_y_0, x = input0_251_cast_fp16, y = value_44_cast_fp16)[name = string("x2_42_cast_fp16")];
            tensor<int32, [4]> var_4624_perm_0 = const()[name = string("op_4624_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4625 = const()[name = string("op_4625"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_4624_cast_fp16 = transpose(perm = var_4624_perm_0, x = x2_42_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [1, 4, 1024]> input1_126_cast_fp16 = reshape(shape = var_4625, x = var_4624_cast_fp16)[name = string("input1_126_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501998848))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503047488))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_palettized, x = input1_126_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_253_cast_fp16 = add(x = input_253_cast_fp16, y = linear_187_cast_fp16)[name = string("input0_253_cast_fp16")];
            tensor<int32, [1]> x_415_axes_0 = const()[name = string("x_415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503048064)))];
            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503050176)))];
            tensor<fp16, [1, 4, 1024]> x_415_cast_fp16 = layer_norm(axes = x_415_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input0_253_cast_fp16)[name = string("x_415_cast_fp16")];
            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_255_pad_type_0 = const()[name = string("input0_255_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_255_strides_0 = const()[name = string("input0_255_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_255_pad_0 = const()[name = string("input0_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_255_dilations_0 = const()[name = string("input0_255_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_255_groups_0 = const()[name = string("input0_255_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503052288))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505149504))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_415_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 2048, 4]> input0_255_cast_fp16 = conv(dilations = input0_255_dilations_0, groups = input0_255_groups_0, pad = input0_255_pad_0, pad_type = input0_255_pad_type_0, strides = input0_255_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_257_cast_fp16)[name = string("input0_255_cast_fp16")];
            int32 x_417_split_num_splits_0 = const()[name = string("x_417_split_num_splits_0"), val = int32(2)];
            int32 x_417_split_axis_0 = const()[name = string("x_417_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_417_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_417_split_cast_fp16_1 = split(axis = x_417_split_axis_0, num_splits = x_417_split_num_splits_0, x = input0_255_cast_fp16)[name = string("x_417_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_417_split_1_sigmoid_cast_fp16 = sigmoid(x = x_417_split_cast_fp16_1)[name = string("x_417_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_417_cast_fp16 = mul(x = x_417_split_cast_fp16_0, y = x_417_split_1_sigmoid_cast_fp16)[name = string("x_417_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_257_cast_fp16 = select(a = var_30_to_fp16, b = x_417_cast_fp16, cond = var_570)[name = string("input0_257_cast_fp16")];
            bool new_x0_42_interleave_0 = const()[name = string("new_x0_42_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_42_cast_fp16 = concat(axis = var_56, interleave = new_x0_42_interleave_0, values = (cache40_1_cast_fp16, input0_257_cast_fp16))[name = string("new_x0_42_cast_fp16")];
            tensor<int32, [3]> var_4663_begin_0 = const()[name = string("op_4663_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4663_end_0 = const()[name = string("op_4663_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_4663_end_mask_0 = const()[name = string("op_4663_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_4663_cast_fp16 = slice_by_index(begin = var_4663_begin_0, end = var_4663_end_0, end_mask = var_4663_end_mask_0, x = new_x0_42_cast_fp16)[name = string("op_4663_cast_fp16")];
            string x_419_pad_type_0 = const()[name = string("x_419_pad_type_0"), val = string("valid")];
            int32 x_419_groups_0 = const()[name = string("x_419_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_419_strides_0 = const()[name = string("x_419_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_419_pad_0 = const()[name = string("x_419_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_419_dilations_0 = const()[name = string("x_419_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_20_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505150080))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505159360))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_419_cast_fp16 = conv(dilations = x_419_dilations_0, groups = x_419_groups_0, pad = x_419_pad_0, pad_type = x_419_pad_type_0, strides = x_419_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_42_cast_fp16)[name = string("x_419_cast_fp16")];
            tensor<int32, [3]> input1_128_perm_0 = const()[name = string("input1_128_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_421_axes_0 = const()[name = string("x_421_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505159936)))];
            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505162048)))];
            tensor<fp16, [1, 4, 1024]> input1_128_cast_fp16 = transpose(perm = input1_128_perm_0, x = x_419_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [1, 4, 1024]> x_421_cast_fp16 = layer_norm(axes = x_421_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input1_128_cast_fp16)[name = string("x_421_cast_fp16")];
            tensor<int32, [3]> input2_84_perm_0 = const()[name = string("input2_84_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_84_cast_fp16 = transpose(perm = input2_84_perm_0, x = x_421_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 1024, 4]> var_4678_cast_fp16 = silu(x = input2_84_cast_fp16)[name = string("op_4678_cast_fp16")];
            string x_423_pad_type_0 = const()[name = string("x_423_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_423_strides_0 = const()[name = string("x_423_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_423_pad_0 = const()[name = string("x_423_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_423_dilations_0 = const()[name = string("x_423_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_423_groups_0 = const()[name = string("x_423_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505164160))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506212800))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_423_cast_fp16 = conv(dilations = x_423_dilations_0, groups = x_423_groups_0, pad = x_423_pad_0, pad_type = x_423_pad_type_0, strides = x_423_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_4678_cast_fp16)[name = string("x_423_cast_fp16")];
            tensor<int32, [3]> input3_44_perm_0 = const()[name = string("input3_44_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_44_cast_fp16 = transpose(perm = input3_44_perm_0, x = x_423_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 1024]> input1_130_cast_fp16 = add(x = input0_253_cast_fp16, y = input3_44_cast_fp16)[name = string("input1_130_cast_fp16")];
            tensor<int32, [1]> input0_259_axes_0 = const()[name = string("input0_259_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506213376)))];
            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506215488)))];
            tensor<fp16, [1, 4, 1024]> input0_259_cast_fp16 = layer_norm(axes = input0_259_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input1_130_cast_fp16)[name = string("input0_259_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506217600))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510411968))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_259_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4699_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510412544))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514606912))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_palettized, x = var_4699_cast_fp16)[name = string("linear_189_cast_fp16")];
            fp16 var_4704_to_fp16 = const()[name = string("op_4704_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4705_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4704_to_fp16)[name = string("op_4705_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_86_cast_fp16 = add(x = input1_130_cast_fp16, y = var_4705_cast_fp16)[name = string("input2_86_cast_fp16")];
            tensor<int32, [1]> input0_261_axes_0 = const()[name = string("input0_261_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514607488)))];
            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514609600)))];
            tensor<fp16, [1, 4, 1024]> input0_261_cast_fp16 = layer_norm(axes = input0_261_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input2_86_cast_fp16)[name = string("input0_261_cast_fp16")];
            tensor<int32, [4]> cache41_1_begin_0 = const()[name = string("cache41_1_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> cache41_1_end_0 = const()[name = string("cache41_1_end_0"), val = tensor<int32, [4]>([22, 1, 56, 1024])];
            tensor<bool, [4]> cache41_1_end_mask_0 = const()[name = string("cache41_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache41_1_squeeze_mask_0 = const()[name = string("cache41_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache41_1_cast_fp16 = slice_by_index(begin = cache41_1_begin_0, end = cache41_1_end_0, end_mask = cache41_1_end_mask_0, squeeze_mask = cache41_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache41_1_cast_fp16")];
            tensor<int32, [4]> cache42_1_begin_0 = const()[name = string("cache42_1_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> cache42_1_end_0 = const()[name = string("cache42_1_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
            tensor<bool, [4]> cache42_1_end_mask_0 = const()[name = string("cache42_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache42_1_squeeze_mask_0 = const()[name = string("cache42_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache42_1_cast_fp16 = slice_by_index(begin = cache42_1_begin_0, end = cache42_1_end_0, end_mask = cache42_1_end_mask_0, squeeze_mask = cache42_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache42_1_cast_fp16")];
            tensor<int32, [1]> input_261_axes_0 = const()[name = string("input_261_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514611712)))];
            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514613824)))];
            tensor<fp16, [1, 4, 1024]> input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input0_261_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514615936))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518810304))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_palettized, x = input_261_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4734_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("op_4734_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518810880))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523005248))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_palettized, x = var_4734_cast_fp16)[name = string("linear_191_cast_fp16")];
            fp16 var_4739_to_fp16 = const()[name = string("op_4739_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4740_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4739_to_fp16)[name = string("op_4740_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_265_cast_fp16 = add(x = input0_261_cast_fp16, y = var_4740_cast_fp16)[name = string("input_265_cast_fp16")];
            tensor<int32, [1]> key_44_axes_0 = const()[name = string("key_44_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523005824)))];
            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523007936)))];
            tensor<fp16, [1, 4, 1024]> key_44_cast_fp16 = layer_norm(axes = key_44_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_265_cast_fp16)[name = string("key_44_cast_fp16")];
            bool input_267_interleave_0 = const()[name = string("input_267_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_267_cast_fp16 = concat(axis = var_64, interleave = input_267_interleave_0, values = (cache41_1_cast_fp16, key_44_cast_fp16))[name = string("input_267_cast_fp16")];
            tensor<int32, [3]> var_4762_begin_0 = const()[name = string("op_4762_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_4762_end_0 = const()[name = string("op_4762_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_4762_end_mask_0 = const()[name = string("op_4762_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_4762_cast_fp16 = slice_by_index(begin = var_4762_begin_0, end = var_4762_end_0, end_mask = var_4762_end_mask_0, x = cache41_1_cast_fp16)[name = string("op_4762_cast_fp16")];
            bool var_4768_interleave_0 = const()[name = string("op_4768_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_4768_cast_fp16 = concat(axis = var_64, interleave = var_4768_interleave_0, values = (var_4762_cast_fp16, key_44_cast_fp16))[name = string("op_4768_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523010048))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524058688))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_palettized, x = key_44_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_4772 = const()[name = string("op_4772"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_44_cast_fp16 = reshape(shape = var_4772, x = linear_192_cast_fp16)[name = string("q_44_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524059264))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525107904))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<int32, [4]> var_4776 = const()[name = string("op_4776"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_44_cast_fp16 = reshape(shape = var_4776, x = linear_193_cast_fp16)[name = string("k_44_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525108480))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526157120))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_194_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<int32, [4]> var_4780 = const()[name = string("op_4780"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_44_cast_fp16 = reshape(shape = var_4780, x = linear_194_cast_fp16)[name = string("v_44_cast_fp16")];
            tensor<int32, [4]> value_46_perm_0 = const()[name = string("value_46_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526157696)))];
            tensor<fp16, [1, 4, 8, 128]> var_4792_cast_fp16 = add(x = q_44_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4792_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526159808)))];
            tensor<fp16, [1, 4, 8, 128]> var_4794_cast_fp16 = add(x = q_44_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4794_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_44_perm_0 = const()[name = string("q_with_bias_v_44_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_431_transpose_x_0 = const()[name = string("x_431_transpose_x_0"), val = bool(false)];
            bool x_431_transpose_y_0 = const()[name = string("x_431_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_4796_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526161920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526283840))))[name = string("op_4796_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_44_cast_fp16 = transpose(perm = q_with_bias_v_44_perm_0, x = var_4794_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 8, 4, 119]> x_431_cast_fp16 = matmul(transpose_x = x_431_transpose_x_0, transpose_y = x_431_transpose_y_0, x = q_with_bias_v_44_cast_fp16, y = op_4796_to_fp16_palettized)[name = string("x_431_cast_fp16")];
            tensor<int32, [8]> x0_46_pad_0 = const()[name = string("x0_46_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_46_mode_0 = const()[name = string("x0_46_mode_0"), val = string("constant")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_46_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x0_46_mode_0, pad = x0_46_pad_0, x = x_431_cast_fp16)[name = string("x0_46_cast_fp16")];
            tensor<int32, [4]> var_4804 = const()[name = string("op_4804"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_44_cast_fp16 = reshape(shape = var_4804, x = x0_46_cast_fp16)[name = string("x1_44_cast_fp16")];
            tensor<int32, [4]> var_4808_begin_0 = const()[name = string("op_4808_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4808_end_0 = const()[name = string("op_4808_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_4808_end_mask_0 = const()[name = string("op_4808_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_4808_cast_fp16 = slice_by_index(begin = var_4808_begin_0, end = var_4808_end_0, end_mask = var_4808_end_mask_0, x = x1_44_cast_fp16)[name = string("op_4808_cast_fp16")];
            tensor<int32, [4]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_44_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("matrix_bd_44_cast_fp16")];
            bool matrix_ac_44_transpose_x_0 = const()[name = string("matrix_ac_44_transpose_x_0"), val = bool(false)];
            bool matrix_ac_44_transpose_y_0 = const()[name = string("matrix_ac_44_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_44_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 8, 4, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4792_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_44_cast_fp16 = matmul(transpose_x = matrix_ac_44_transpose_x_0, transpose_y = matrix_ac_44_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_44_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_44_begin_0 = const()[name = string("matrix_bd0_44_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_44_end_0 = const()[name = string("matrix_bd0_44_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_44_end_mask_0 = const()[name = string("matrix_bd0_44_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_44_cast_fp16 = slice_by_index(begin = matrix_bd0_44_begin_0, end = matrix_bd0_44_end_0, end_mask = matrix_bd0_44_end_mask_0, x = matrix_bd_44_cast_fp16)[name = string("matrix_bd0_44_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4818_cast_fp16 = add(x = matrix_ac_44_cast_fp16, y = matrix_bd0_44_cast_fp16)[name = string("op_4818_cast_fp16")];
            fp16 _inversed_scores_44_y_0_to_fp16 = const()[name = string("_inversed_scores_44_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_44_cast_fp16 = mul(x = var_4818_cast_fp16, y = _inversed_scores_44_y_0_to_fp16)[name = string("_inversed_scores_44_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_44_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_44_cast_fp16, cond = mask0_4)[name = string("scores0_44_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_4824_cast_fp16 = softmax(axis = var_56, x = scores0_44_cast_fp16)[name = string("op_4824_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_263_cast_fp16 = select(a = var_30_to_fp16, b = var_4824_cast_fp16, cond = mask0_4)[name = string("input0_263_cast_fp16")];
            bool x2_44_transpose_x_0 = const()[name = string("x2_44_transpose_x_0"), val = bool(false)];
            bool x2_44_transpose_y_0 = const()[name = string("x2_44_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_46_cast_fp16 = transpose(perm = value_46_perm_0, x = v_44_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 8, 4, 128]> x2_44_cast_fp16 = matmul(transpose_x = x2_44_transpose_x_0, transpose_y = x2_44_transpose_y_0, x = input0_263_cast_fp16, y = value_46_cast_fp16)[name = string("x2_44_cast_fp16")];
            tensor<int32, [4]> var_4828_perm_0 = const()[name = string("op_4828_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4829 = const()[name = string("op_4829"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_4828_cast_fp16 = transpose(perm = var_4828_perm_0, x = x2_44_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [1, 4, 1024]> input1_132_cast_fp16 = reshape(shape = var_4829, x = var_4828_cast_fp16)[name = string("input1_132_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526284416))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527333056))))[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16_palettized, x = input1_132_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_265_cast_fp16 = add(x = input_265_cast_fp16, y = linear_196_cast_fp16)[name = string("input0_265_cast_fp16")];
            tensor<int32, [1]> x_435_axes_0 = const()[name = string("x_435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527333632)))];
            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527335744)))];
            tensor<fp16, [1, 4, 1024]> x_435_cast_fp16 = layer_norm(axes = x_435_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input0_265_cast_fp16)[name = string("x_435_cast_fp16")];
            tensor<int32, [3]> input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_267_pad_type_0 = const()[name = string("input0_267_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_267_strides_0 = const()[name = string("input0_267_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_267_pad_0 = const()[name = string("input0_267_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_267_dilations_0 = const()[name = string("input0_267_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_267_groups_0 = const()[name = string("input0_267_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527337856))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529435072))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_435_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 2048, 4]> input0_267_cast_fp16 = conv(dilations = input0_267_dilations_0, groups = input0_267_groups_0, pad = input0_267_pad_0, pad_type = input0_267_pad_type_0, strides = input0_267_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_269_cast_fp16)[name = string("input0_267_cast_fp16")];
            int32 x_437_split_num_splits_0 = const()[name = string("x_437_split_num_splits_0"), val = int32(2)];
            int32 x_437_split_axis_0 = const()[name = string("x_437_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_437_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_437_split_cast_fp16_1 = split(axis = x_437_split_axis_0, num_splits = x_437_split_num_splits_0, x = input0_267_cast_fp16)[name = string("x_437_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_437_split_1_sigmoid_cast_fp16 = sigmoid(x = x_437_split_cast_fp16_1)[name = string("x_437_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_437_cast_fp16 = mul(x = x_437_split_cast_fp16_0, y = x_437_split_1_sigmoid_cast_fp16)[name = string("x_437_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_269_cast_fp16 = select(a = var_30_to_fp16, b = x_437_cast_fp16, cond = var_570)[name = string("input0_269_cast_fp16")];
            bool new_x0_44_interleave_0 = const()[name = string("new_x0_44_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_44_cast_fp16 = concat(axis = var_56, interleave = new_x0_44_interleave_0, values = (cache42_1_cast_fp16, input0_269_cast_fp16))[name = string("new_x0_44_cast_fp16")];
            tensor<int32, [3]> var_4867_begin_0 = const()[name = string("op_4867_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4867_end_0 = const()[name = string("op_4867_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_4867_end_mask_0 = const()[name = string("op_4867_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_4867_cast_fp16 = slice_by_index(begin = var_4867_begin_0, end = var_4867_end_0, end_mask = var_4867_end_mask_0, x = new_x0_44_cast_fp16)[name = string("op_4867_cast_fp16")];
            string x_439_pad_type_0 = const()[name = string("x_439_pad_type_0"), val = string("valid")];
            int32 x_439_groups_0 = const()[name = string("x_439_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_439_strides_0 = const()[name = string("x_439_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_439_pad_0 = const()[name = string("x_439_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_439_dilations_0 = const()[name = string("x_439_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_21_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529435648))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529444928))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_439_cast_fp16 = conv(dilations = x_439_dilations_0, groups = x_439_groups_0, pad = x_439_pad_0, pad_type = x_439_pad_type_0, strides = x_439_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_44_cast_fp16)[name = string("x_439_cast_fp16")];
            tensor<int32, [3]> input1_134_perm_0 = const()[name = string("input1_134_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_441_axes_0 = const()[name = string("x_441_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529445504)))];
            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529447616)))];
            tensor<fp16, [1, 4, 1024]> input1_134_cast_fp16 = transpose(perm = input1_134_perm_0, x = x_439_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [1, 4, 1024]> x_441_cast_fp16 = layer_norm(axes = x_441_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input1_134_cast_fp16)[name = string("x_441_cast_fp16")];
            tensor<int32, [3]> input2_88_perm_0 = const()[name = string("input2_88_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_88_cast_fp16 = transpose(perm = input2_88_perm_0, x = x_441_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 1024, 4]> var_4882_cast_fp16 = silu(x = input2_88_cast_fp16)[name = string("op_4882_cast_fp16")];
            string x_443_pad_type_0 = const()[name = string("x_443_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_443_strides_0 = const()[name = string("x_443_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_443_pad_0 = const()[name = string("x_443_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_443_dilations_0 = const()[name = string("x_443_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_443_groups_0 = const()[name = string("x_443_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529449728))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530498368))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_443_cast_fp16 = conv(dilations = x_443_dilations_0, groups = x_443_groups_0, pad = x_443_pad_0, pad_type = x_443_pad_type_0, strides = x_443_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_4882_cast_fp16)[name = string("x_443_cast_fp16")];
            tensor<int32, [3]> input3_46_perm_0 = const()[name = string("input3_46_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_46_cast_fp16 = transpose(perm = input3_46_perm_0, x = x_443_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 1024]> input1_136_cast_fp16 = add(x = input0_265_cast_fp16, y = input3_46_cast_fp16)[name = string("input1_136_cast_fp16")];
            tensor<int32, [1]> input0_271_axes_0 = const()[name = string("input0_271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530498944)))];
            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530501056)))];
            tensor<fp16, [1, 4, 1024]> input0_271_cast_fp16 = layer_norm(axes = input0_271_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input1_136_cast_fp16)[name = string("input0_271_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530503168))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534697536))))[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_197_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_271_cast_fp16)[name = string("linear_197_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4903_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("op_4903_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534698112))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538892480))))[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16_palettized, x = var_4903_cast_fp16)[name = string("linear_198_cast_fp16")];
            fp16 var_4908_to_fp16 = const()[name = string("op_4908_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4909_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_4908_to_fp16)[name = string("op_4909_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_90_cast_fp16 = add(x = input1_136_cast_fp16, y = var_4909_cast_fp16)[name = string("input2_90_cast_fp16")];
            tensor<int32, [1]> input0_273_axes_0 = const()[name = string("input0_273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538893056)))];
            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538895168)))];
            tensor<fp16, [1, 4, 1024]> input0_273_cast_fp16 = layer_norm(axes = input0_273_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input2_90_cast_fp16)[name = string("input0_273_cast_fp16")];
            tensor<int32, [4]> cache43_1_begin_0 = const()[name = string("cache43_1_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> cache43_1_end_0 = const()[name = string("cache43_1_end_0"), val = tensor<int32, [4]>([23, 1, 56, 1024])];
            tensor<bool, [4]> cache43_1_end_mask_0 = const()[name = string("cache43_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache43_1_squeeze_mask_0 = const()[name = string("cache43_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache43_1_cast_fp16 = slice_by_index(begin = cache43_1_begin_0, end = cache43_1_end_0, end_mask = cache43_1_end_mask_0, squeeze_mask = cache43_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache43_1_cast_fp16")];
            tensor<int32, [4]> cache44_1_begin_0 = const()[name = string("cache44_1_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> cache44_1_end_0 = const()[name = string("cache44_1_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
            tensor<bool, [4]> cache44_1_end_mask_0 = const()[name = string("cache44_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache44_1_squeeze_mask_0 = const()[name = string("cache44_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache44_1_cast_fp16 = slice_by_index(begin = cache44_1_begin_0, end = cache44_1_end_0, end_mask = cache44_1_end_mask_0, squeeze_mask = cache44_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache44_1_cast_fp16")];
            tensor<int32, [1]> input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538897280)))];
            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538899392)))];
            tensor<fp16, [1, 4, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input0_273_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538901504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543095872))))[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_4938_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("op_4938_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543096448))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547290816))))[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16_palettized, x = var_4938_cast_fp16)[name = string("linear_200_cast_fp16")];
            fp16 var_4943_to_fp16 = const()[name = string("op_4943_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_4944_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_4943_to_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_277_cast_fp16 = add(x = input0_273_cast_fp16, y = var_4944_cast_fp16)[name = string("input_277_cast_fp16")];
            tensor<int32, [1]> key_46_axes_0 = const()[name = string("key_46_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547291392)))];
            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547293504)))];
            tensor<fp16, [1, 4, 1024]> key_46_cast_fp16 = layer_norm(axes = key_46_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_277_cast_fp16)[name = string("key_46_cast_fp16")];
            bool input_279_interleave_0 = const()[name = string("input_279_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_279_cast_fp16 = concat(axis = var_64, interleave = input_279_interleave_0, values = (cache43_1_cast_fp16, key_46_cast_fp16))[name = string("input_279_cast_fp16")];
            tensor<int32, [3]> var_4966_begin_0 = const()[name = string("op_4966_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_4966_end_0 = const()[name = string("op_4966_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_4966_end_mask_0 = const()[name = string("op_4966_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_4966_cast_fp16 = slice_by_index(begin = var_4966_begin_0, end = var_4966_end_0, end_mask = var_4966_end_mask_0, x = cache43_1_cast_fp16)[name = string("op_4966_cast_fp16")];
            bool var_4972_interleave_0 = const()[name = string("op_4972_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> var_4972_cast_fp16 = concat(axis = var_64, interleave = var_4972_interleave_0, values = (var_4966_cast_fp16, key_46_cast_fp16))[name = string("op_4972_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547295616))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548344256))))[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16_palettized, x = key_46_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_4976 = const()[name = string("op_4976"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_46_cast_fp16 = reshape(shape = var_4976, x = linear_201_cast_fp16)[name = string("q_46_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548344832))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549393472))))[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16_palettized, x = input_279_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<int32, [4]> var_4980 = const()[name = string("op_4980"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_46_cast_fp16 = reshape(shape = var_4980, x = linear_202_cast_fp16)[name = string("k_46_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549394048))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550442688))))[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_203_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16_palettized, x = input_279_cast_fp16)[name = string("linear_203_cast_fp16")];
            tensor<int32, [4]> var_4984 = const()[name = string("op_4984"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_46_cast_fp16 = reshape(shape = var_4984, x = linear_203_cast_fp16)[name = string("v_46_cast_fp16")];
            tensor<int32, [4]> value_48_perm_0 = const()[name = string("value_48_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550443264)))];
            tensor<fp16, [1, 4, 8, 128]> var_4996_cast_fp16 = add(x = q_46_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_4996_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550445376)))];
            tensor<fp16, [1, 4, 8, 128]> var_4998_cast_fp16 = add(x = q_46_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_4998_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_46_perm_0 = const()[name = string("q_with_bias_v_46_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_451_transpose_x_0 = const()[name = string("x_451_transpose_x_0"), val = bool(false)];
            bool x_451_transpose_y_0 = const()[name = string("x_451_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_5000_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550447488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550569408))))[name = string("op_5000_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_46_cast_fp16 = transpose(perm = q_with_bias_v_46_perm_0, x = var_4998_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 8, 4, 119]> x_451_cast_fp16 = matmul(transpose_x = x_451_transpose_x_0, transpose_y = x_451_transpose_y_0, x = q_with_bias_v_46_cast_fp16, y = op_5000_to_fp16_palettized)[name = string("x_451_cast_fp16")];
            tensor<int32, [8]> x0_48_pad_0 = const()[name = string("x0_48_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_48_mode_0 = const()[name = string("x0_48_mode_0"), val = string("constant")];
            fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_48_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x0_48_mode_0, pad = x0_48_pad_0, x = x_451_cast_fp16)[name = string("x0_48_cast_fp16")];
            tensor<int32, [4]> var_5008 = const()[name = string("op_5008"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_46_cast_fp16 = reshape(shape = var_5008, x = x0_48_cast_fp16)[name = string("x1_46_cast_fp16")];
            tensor<int32, [4]> var_5012_begin_0 = const()[name = string("op_5012_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5012_end_0 = const()[name = string("op_5012_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_5012_end_mask_0 = const()[name = string("op_5012_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_5012_cast_fp16 = slice_by_index(begin = var_5012_begin_0, end = var_5012_end_0, end_mask = var_5012_end_mask_0, x = x1_46_cast_fp16)[name = string("op_5012_cast_fp16")];
            tensor<int32, [4]> var_5013 = const()[name = string("op_5013"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_46_cast_fp16 = reshape(shape = var_5013, x = var_5012_cast_fp16)[name = string("matrix_bd_46_cast_fp16")];
            bool matrix_ac_46_transpose_x_0 = const()[name = string("matrix_ac_46_transpose_x_0"), val = bool(false)];
            bool matrix_ac_46_transpose_y_0 = const()[name = string("matrix_ac_46_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_46_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 8, 4, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_4996_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_46_cast_fp16 = matmul(transpose_x = matrix_ac_46_transpose_x_0, transpose_y = matrix_ac_46_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_46_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_46_begin_0 = const()[name = string("matrix_bd0_46_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_46_end_0 = const()[name = string("matrix_bd0_46_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_46_end_mask_0 = const()[name = string("matrix_bd0_46_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_46_cast_fp16 = slice_by_index(begin = matrix_bd0_46_begin_0, end = matrix_bd0_46_end_0, end_mask = matrix_bd0_46_end_mask_0, x = matrix_bd_46_cast_fp16)[name = string("matrix_bd0_46_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_5022_cast_fp16 = add(x = matrix_ac_46_cast_fp16, y = matrix_bd0_46_cast_fp16)[name = string("op_5022_cast_fp16")];
            fp16 _inversed_scores_46_y_0_to_fp16 = const()[name = string("_inversed_scores_46_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_46_cast_fp16 = mul(x = var_5022_cast_fp16, y = _inversed_scores_46_y_0_to_fp16)[name = string("_inversed_scores_46_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_46_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_46_cast_fp16, cond = mask0_4)[name = string("scores0_46_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_5028_cast_fp16 = softmax(axis = var_56, x = scores0_46_cast_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_275_cast_fp16 = select(a = var_30_to_fp16, b = var_5028_cast_fp16, cond = mask0_4)[name = string("input0_275_cast_fp16")];
            bool x2_46_transpose_x_0 = const()[name = string("x2_46_transpose_x_0"), val = bool(false)];
            bool x2_46_transpose_y_0 = const()[name = string("x2_46_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_48_cast_fp16 = transpose(perm = value_48_perm_0, x = v_46_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 8, 4, 128]> x2_46_cast_fp16 = matmul(transpose_x = x2_46_transpose_x_0, transpose_y = x2_46_transpose_y_0, x = input0_275_cast_fp16, y = value_48_cast_fp16)[name = string("x2_46_cast_fp16")];
            tensor<int32, [4]> var_5032_perm_0 = const()[name = string("op_5032_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5033 = const()[name = string("op_5033"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_5032_cast_fp16 = transpose(perm = var_5032_perm_0, x = x2_46_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 4, 1024]> input1_138_cast_fp16 = reshape(shape = var_5033, x = var_5032_cast_fp16)[name = string("input1_138_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550569984))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551618624))))[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16_palettized, x = input1_138_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_277_cast_fp16 = add(x = input_277_cast_fp16, y = linear_205_cast_fp16)[name = string("input0_277_cast_fp16")];
            tensor<int32, [1]> x_455_axes_0 = const()[name = string("x_455_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551619200)))];
            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551621312)))];
            tensor<fp16, [1, 4, 1024]> x_455_cast_fp16 = layer_norm(axes = x_455_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input0_277_cast_fp16)[name = string("x_455_cast_fp16")];
            tensor<int32, [3]> input_281_perm_0 = const()[name = string("input_281_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_279_pad_type_0 = const()[name = string("input0_279_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_279_strides_0 = const()[name = string("input0_279_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_279_pad_0 = const()[name = string("input0_279_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_279_dilations_0 = const()[name = string("input0_279_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_279_groups_0 = const()[name = string("input0_279_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551623424))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553720640))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_281_cast_fp16 = transpose(perm = input_281_perm_0, x = x_455_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2048, 4]> input0_279_cast_fp16 = conv(dilations = input0_279_dilations_0, groups = input0_279_groups_0, pad = input0_279_pad_0, pad_type = input0_279_pad_type_0, strides = input0_279_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_281_cast_fp16)[name = string("input0_279_cast_fp16")];
            int32 x_457_split_num_splits_0 = const()[name = string("x_457_split_num_splits_0"), val = int32(2)];
            int32 x_457_split_axis_0 = const()[name = string("x_457_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_457_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_457_split_cast_fp16_1 = split(axis = x_457_split_axis_0, num_splits = x_457_split_num_splits_0, x = input0_279_cast_fp16)[name = string("x_457_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_457_split_1_sigmoid_cast_fp16 = sigmoid(x = x_457_split_cast_fp16_1)[name = string("x_457_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_457_cast_fp16 = mul(x = x_457_split_cast_fp16_0, y = x_457_split_1_sigmoid_cast_fp16)[name = string("x_457_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_281_cast_fp16 = select(a = var_30_to_fp16, b = x_457_cast_fp16, cond = var_570)[name = string("input0_281_cast_fp16")];
            bool new_x0_46_interleave_0 = const()[name = string("new_x0_46_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_46_cast_fp16 = concat(axis = var_56, interleave = new_x0_46_interleave_0, values = (cache44_1_cast_fp16, input0_281_cast_fp16))[name = string("new_x0_46_cast_fp16")];
            tensor<int32, [3]> var_5071_begin_0 = const()[name = string("op_5071_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_5071_end_0 = const()[name = string("op_5071_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> var_5071_end_mask_0 = const()[name = string("op_5071_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> var_5071_cast_fp16 = slice_by_index(begin = var_5071_begin_0, end = var_5071_end_0, end_mask = var_5071_end_mask_0, x = new_x0_46_cast_fp16)[name = string("op_5071_cast_fp16")];
            string x_459_pad_type_0 = const()[name = string("x_459_pad_type_0"), val = string("valid")];
            int32 x_459_groups_0 = const()[name = string("x_459_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_459_strides_0 = const()[name = string("x_459_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_459_pad_0 = const()[name = string("x_459_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_459_dilations_0 = const()[name = string("x_459_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_22_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553721216))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553730496))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_459_cast_fp16 = conv(dilations = x_459_dilations_0, groups = x_459_groups_0, pad = x_459_pad_0, pad_type = x_459_pad_type_0, strides = x_459_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_46_cast_fp16)[name = string("x_459_cast_fp16")];
            tensor<int32, [3]> input1_140_perm_0 = const()[name = string("input1_140_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_461_axes_0 = const()[name = string("x_461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553731072)))];
            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553733184)))];
            tensor<fp16, [1, 4, 1024]> input1_140_cast_fp16 = transpose(perm = input1_140_perm_0, x = x_459_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 4, 1024]> x_461_cast_fp16 = layer_norm(axes = x_461_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input1_140_cast_fp16)[name = string("x_461_cast_fp16")];
            tensor<int32, [3]> input2_92_perm_0 = const()[name = string("input2_92_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_92_cast_fp16 = transpose(perm = input2_92_perm_0, x = x_461_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 1024, 4]> var_5086_cast_fp16 = silu(x = input2_92_cast_fp16)[name = string("op_5086_cast_fp16")];
            string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553735296))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554783936))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_5086_cast_fp16)[name = string("x_463_cast_fp16")];
            tensor<int32, [3]> input3_48_perm_0 = const()[name = string("input3_48_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_48_cast_fp16 = transpose(perm = input3_48_perm_0, x = x_463_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 4, 1024]> input1_142_cast_fp16 = add(x = input0_277_cast_fp16, y = input3_48_cast_fp16)[name = string("input1_142_cast_fp16")];
            tensor<int32, [1]> input0_10_axes_0 = const()[name = string("input0_10_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554784512)))];
            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554786624)))];
            tensor<fp16, [1, 4, 1024]> input0_10_cast_fp16 = layer_norm(axes = input0_10_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input1_142_cast_fp16)[name = string("input0_10_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554788736))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558983104))))[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_206_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_10_cast_fp16)[name = string("linear_206_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_5107_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("op_5107_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558983680))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563178048))))[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16_palettized, x = var_5107_cast_fp16)[name = string("linear_207_cast_fp16")];
            fp16 var_5112_to_fp16 = const()[name = string("op_5112_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_5113_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5112_to_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_94_cast_fp16 = add(x = input1_142_cast_fp16, y = var_5113_cast_fp16)[name = string("input2_94_cast_fp16")];
            tensor<int32, [1]> input0_283_axes_0 = const()[name = string("input0_283_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563178624)))];
            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563180736)))];
            tensor<fp16, [1, 4, 1024]> input0_283_cast_fp16 = layer_norm(axes = input0_283_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input2_94_cast_fp16)[name = string("input0_283_cast_fp16")];
            tensor<int32, [4]> cache45_1_begin_0 = const()[name = string("cache45_1_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> cache45_1_end_0 = const()[name = string("cache45_1_end_0"), val = tensor<int32, [4]>([24, 1, 56, 1024])];
            tensor<bool, [4]> cache45_1_end_mask_0 = const()[name = string("cache45_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache45_1_squeeze_mask_0 = const()[name = string("cache45_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 56, 1024]> cache45_1_cast_fp16 = slice_by_index(begin = cache45_1_begin_0, end = cache45_1_end_0, end_mask = cache45_1_end_mask_0, squeeze_mask = cache45_1_squeeze_mask_0, x = cache_last_channel_to_fp16)[name = string("cache45_1_cast_fp16")];
            tensor<int32, [4]> cache46_1_begin_0 = const()[name = string("cache46_1_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> cache46_1_end_0 = const()[name = string("cache46_1_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
            tensor<bool, [4]> cache46_1_end_mask_0 = const()[name = string("cache46_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> cache46_1_squeeze_mask_0 = const()[name = string("cache46_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 1024, 8]> cache46_1_cast_fp16 = slice_by_index(begin = cache46_1_begin_0, end = cache46_1_end_0, end_mask = cache46_1_end_mask_0, squeeze_mask = cache46_1_squeeze_mask_0, x = cache_last_time_to_fp16)[name = string("cache46_1_cast_fp16")];
            tensor<int32, [1]> input_4_axes_0 = const()[name = string("input_4_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563182848)))];
            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563184960)))];
            tensor<fp16, [1, 4, 1024]> input_4_cast_fp16 = layer_norm(axes = input_4_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input0_283_cast_fp16)[name = string("input_4_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563187072))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567381440))))[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_208_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16_palettized, x = input_4_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_5142_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("op_5142_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567382016))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571576384))))[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16_palettized, x = var_5142_cast_fp16)[name = string("linear_209_cast_fp16")];
            fp16 var_5147_to_fp16 = const()[name = string("op_5147_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_5148_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5147_to_fp16)[name = string("op_5148_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input_2_cast_fp16 = add(x = input0_283_cast_fp16, y = var_5148_cast_fp16)[name = string("input_2_cast_fp16")];
            tensor<int32, [1]> key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571576960)))];
            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571579072)))];
            tensor<fp16, [1, 4, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_2_cast_fp16)[name = string("key_1_cast_fp16")];
            bool input_8_interleave_0 = const()[name = string("input_8_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 60, 1024]> input_8_cast_fp16 = concat(axis = var_64, interleave = input_8_interleave_0, values = (cache45_1_cast_fp16, key_1_cast_fp16))[name = string("input_8_cast_fp16")];
            tensor<int32, [3]> var_5170_begin_0 = const()[name = string("op_5170_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_5170_end_0 = const()[name = string("op_5170_end_0"), val = tensor<int32, [3]>([1, 56, 1024])];
            tensor<bool, [3]> var_5170_end_mask_0 = const()[name = string("op_5170_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 52, 1024]> var_5170_cast_fp16 = slice_by_index(begin = var_5170_begin_0, end = var_5170_end_0, end_mask = var_5170_end_mask_0, x = cache45_1_cast_fp16)[name = string("op_5170_cast_fp16")];
            bool cache_last_channel_cur_1_interleave_0 = const()[name = string("cache_last_channel_cur_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 56, 1024]> cache_last_channel_cur_1_cast_fp16 = concat(axis = var_64, interleave = cache_last_channel_cur_1_interleave_0, values = (var_5170_cast_fp16, key_1_cast_fp16))[name = string("cache_last_channel_cur_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571581184))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572629824))))[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16_palettized, x = key_1_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_5180 = const()[name = string("op_5180"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 4, 8, 128]> q_1_cast_fp16 = reshape(shape = var_5180, x = linear_210_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572630400))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573679040))))[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16_palettized, x = input_8_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<int32, [4]> var_5184 = const()[name = string("op_5184"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> k_1_cast_fp16 = reshape(shape = var_5184, x = linear_211_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573679616))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574728256))))[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16_palettized")];
            tensor<fp16, [1, 60, 1024]> linear_212_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16_palettized, x = input_8_cast_fp16)[name = string("linear_212_cast_fp16")];
            tensor<int32, [4]> var_5188 = const()[name = string("op_5188"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 60, 8, 128]> v_1_cast_fp16 = reshape(shape = var_5188, x = linear_212_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<int32, [4]> value_1_perm_0 = const()[name = string("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574728832)))];
            tensor<fp16, [1, 4, 8, 128]> var_5200_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5200_cast_fp16")];
            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574730944)))];
            tensor<fp16, [1, 4, 8, 128]> var_5202_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5202_cast_fp16")];
            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool x_10_transpose_x_0 = const()[name = string("x_10_transpose_x_0"), val = bool(false)];
            bool x_10_transpose_y_0 = const()[name = string("x_10_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 119]> op_5204_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 8, 128, 119]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574733056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574854976))))[name = string("op_5204_to_fp16_palettized")];
            tensor<fp16, [1, 8, 4, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_5202_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 8, 4, 119]> x_10_cast_fp16 = matmul(transpose_x = x_10_transpose_x_0, transpose_y = x_10_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_5204_to_fp16_palettized)[name = string("x_10_cast_fp16")];
            tensor<int32, [8]> x0_1_pad_0 = const()[name = string("x0_1_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x0_1_mode_0 = const()[name = string("x0_1_mode_0"), val = string("constant")];
            fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 4, 120]> x0_1_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x0_1_mode_0, pad = x0_1_pad_0, x = x_10_cast_fp16)[name = string("x0_1_cast_fp16")];
            tensor<int32, [4]> var_5212 = const()[name = string("op_5212"), val = tensor<int32, [4]>([1, 8, -1, 4])];
            tensor<fp16, [1, 8, 120, 4]> x1_1_cast_fp16 = reshape(shape = var_5212, x = x0_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> var_5216_begin_0 = const()[name = string("op_5216_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5216_end_0 = const()[name = string("op_5216_end_0"), val = tensor<int32, [4]>([1, 8, 120, 4])];
            tensor<bool, [4]> var_5216_end_mask_0 = const()[name = string("op_5216_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 119, 4]> var_5216_cast_fp16 = slice_by_index(begin = var_5216_begin_0, end = var_5216_end_0, end_mask = var_5216_end_mask_0, x = x1_1_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [4]> var_5217 = const()[name = string("op_5217"), val = tensor<int32, [4]>([1, 8, 4, 119])];
            tensor<fp16, [1, 8, 4, 119]> matrix_bd_1_cast_fp16 = reshape(shape = var_5217, x = var_5216_cast_fp16)[name = string("matrix_bd_1_cast_fp16")];
            bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
            bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 60]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_1_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 8, 4, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5200_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 8, 4, 60]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_1_cast_fp16")];
            tensor<int32, [4]> matrix_bd0_1_begin_0 = const()[name = string("matrix_bd0_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> matrix_bd0_1_end_0 = const()[name = string("matrix_bd0_1_end_0"), val = tensor<int32, [4]>([1, 8, 4, 60])];
            tensor<bool, [4]> matrix_bd0_1_end_mask_0 = const()[name = string("matrix_bd0_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 4, 60]> matrix_bd0_1_cast_fp16 = slice_by_index(begin = matrix_bd0_1_begin_0, end = matrix_bd0_1_end_0, end_mask = matrix_bd0_1_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd0_1_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_5226_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd0_1_cast_fp16)[name = string("op_5226_cast_fp16")];
            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 4, 60]> _inversed_scores_1_cast_fp16 = mul(x = var_5226_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> scores0_1_cast_fp16 = select(a = var_31_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask0_4)[name = string("scores0_1_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> var_5232_cast_fp16 = softmax(axis = var_56, x = scores0_1_cast_fp16)[name = string("op_5232_cast_fp16")];
            tensor<fp16, [1, 8, 4, 60]> input0_4_cast_fp16 = select(a = var_30_to_fp16, b = var_5232_cast_fp16, cond = mask0_4)[name = string("input0_4_cast_fp16")];
            bool x2_1_transpose_x_0 = const()[name = string("x2_1_transpose_x_0"), val = bool(false)];
            bool x2_1_transpose_y_0 = const()[name = string("x2_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 60, 128]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = v_1_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 8, 4, 128]> x2_1_cast_fp16 = matmul(transpose_x = x2_1_transpose_x_0, transpose_y = x2_1_transpose_y_0, x = input0_4_cast_fp16, y = value_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            tensor<int32, [4]> var_5236_perm_0 = const()[name = string("op_5236_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5237 = const()[name = string("op_5237"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 4, 8, 128]> var_5236_cast_fp16 = transpose(perm = var_5236_perm_0, x = x2_1_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 4, 1024]> input1_4_cast_fp16 = reshape(shape = var_5237, x = var_5236_cast_fp16)[name = string("input1_4_cast_fp16")];
            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574855552))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575904192))))[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16_palettized, x = input1_4_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input0_6_cast_fp16 = add(x = input_2_cast_fp16, y = linear_214_cast_fp16)[name = string("input0_6_cast_fp16")];
            tensor<int32, [1]> x_14_axes_0 = const()[name = string("x_14_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575904768)))];
            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575906880)))];
            tensor<fp16, [1, 4, 1024]> x_14_cast_fp16 = layer_norm(axes = x_14_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input0_6_cast_fp16)[name = string("x_14_cast_fp16")];
            tensor<int32, [3]> input_10_perm_0 = const()[name = string("input_10_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input0_8_pad_type_0 = const()[name = string("input0_8_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> input0_8_strides_0 = const()[name = string("input0_8_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> input0_8_pad_0 = const()[name = string("input0_8_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> input0_8_dilations_0 = const()[name = string("input0_8_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 input0_8_groups_0 = const()[name = string("input0_8_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575908992))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578006208))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> input_10_cast_fp16 = transpose(perm = input_10_perm_0, x = x_14_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2048, 4]> input0_8_cast_fp16 = conv(dilations = input0_8_dilations_0, groups = input0_8_groups_0, pad = input0_8_pad_0, pad_type = input0_8_pad_type_0, strides = input0_8_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_10_cast_fp16)[name = string("input0_8_cast_fp16")];
            int32 x_2_split_num_splits_0 = const()[name = string("x_2_split_num_splits_0"), val = int32(2)];
            int32 x_2_split_axis_0 = const()[name = string("x_2_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 4]> x_2_split_cast_fp16_0, tensor<fp16, [1, 1024, 4]> x_2_split_cast_fp16_1 = split(axis = x_2_split_axis_0, num_splits = x_2_split_num_splits_0, x = input0_8_cast_fp16)[name = string("x_2_split_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_2_split_1_sigmoid_cast_fp16 = sigmoid(x = x_2_split_cast_fp16_1)[name = string("x_2_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> x_2_cast_fp16 = mul(x = x_2_split_cast_fp16_0, y = x_2_split_1_sigmoid_cast_fp16)[name = string("x_2_cast_fp16")];
            tensor<fp16, [1, 1024, 4]> input0_2_cast_fp16 = select(a = var_30_to_fp16, b = x_2_cast_fp16, cond = var_570)[name = string("input0_2_cast_fp16")];
            bool new_x0_1_interleave_0 = const()[name = string("new_x0_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 12]> new_x0_1_cast_fp16 = concat(axis = var_56, interleave = new_x0_1_interleave_0, values = (cache46_1_cast_fp16, input0_2_cast_fp16))[name = string("new_x0_1_cast_fp16")];
            tensor<int32, [3]> cache_last_time_cur_1_begin_0 = const()[name = string("cache_last_time_cur_1_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> cache_last_time_cur_1_end_0 = const()[name = string("cache_last_time_cur_1_end_0"), val = tensor<int32, [3]>([1, 1024, 12])];
            tensor<bool, [3]> cache_last_time_cur_1_end_mask_0 = const()[name = string("cache_last_time_cur_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1024, 8]> cache_last_time_cur_1_cast_fp16 = slice_by_index(begin = cache_last_time_cur_1_begin_0, end = cache_last_time_cur_1_end_0, end_mask = cache_last_time_cur_1_end_mask_0, x = new_x0_1_cast_fp16)[name = string("cache_last_time_cur_1_cast_fp16")];
            string x_4_pad_type_0 = const()[name = string("x_4_pad_type_0"), val = string("valid")];
            int32 x_4_groups_0 = const()[name = string("x_4_groups_0"), val = int32(1024)];
            tensor<int32, [1]> x_4_strides_0 = const()[name = string("x_4_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_4_pad_0 = const()[name = string("x_4_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_4_dilations_0 = const()[name = string("x_4_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> encoder_layers_23_conv_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578006784))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578016064))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_4_cast_fp16 = conv(dilations = x_4_dilations_0, groups = x_4_groups_0, pad = x_4_pad_0, pad_type = x_4_pad_type_0, strides = x_4_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_palettized, x = new_x0_1_cast_fp16)[name = string("x_4_cast_fp16")];
            tensor<int32, [3]> input1_1_perm_0 = const()[name = string("input1_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_6_axes_0 = const()[name = string("x_6_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578016640)))];
            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578018752)))];
            tensor<fp16, [1, 4, 1024]> input1_1_cast_fp16 = transpose(perm = input1_1_perm_0, x = x_4_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 4, 1024]> x_6_cast_fp16 = layer_norm(axes = x_6_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input1_1_cast_fp16)[name = string("x_6_cast_fp16")];
            tensor<int32, [3]> input2_1_perm_0 = const()[name = string("input2_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, 4]> input2_1_cast_fp16 = transpose(perm = input2_1_perm_0, x = x_6_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 1024, 4]> var_5290_cast_fp16 = silu(x = input2_1_cast_fp16)[name = string("op_5290_cast_fp16")];
            string x_16_pad_type_0 = const()[name = string("x_16_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> x_16_strides_0 = const()[name = string("x_16_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_16_pad_0 = const()[name = string("x_16_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_16_dilations_0 = const()[name = string("x_16_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 x_16_groups_0 = const()[name = string("x_16_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578020864))), lut = tensor<fp16, [1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579069504))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 4]> x_16_cast_fp16 = conv(dilations = x_16_dilations_0, groups = x_16_groups_0, pad = x_16_pad_0, pad_type = x_16_pad_type_0, strides = x_16_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_palettized, x = var_5290_cast_fp16)[name = string("x_16_cast_fp16")];
            tensor<int32, [3]> input3_1_perm_0 = const()[name = string("input3_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 4, 1024]> input3_1_cast_fp16 = transpose(perm = input3_1_perm_0, x = x_16_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 4, 1024]> input1_2_cast_fp16 = add(x = input0_6_cast_fp16, y = input3_1_cast_fp16)[name = string("input1_2_cast_fp16")];
            tensor<int32, [1]> input0_1_axes_0 = const()[name = string("input0_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579070080)))];
            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579072192)))];
            tensor<fp16, [1, 4, 1024]> input0_1_cast_fp16 = layer_norm(axes = input0_1_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input1_2_cast_fp16)[name = string("input0_1_cast_fp16")];
            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579074304))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583268672))))[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 4096]> linear_215_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16_palettized, x = input0_1_cast_fp16)[name = string("linear_215_cast_fp16")];
            tensor<fp16, [1, 4, 4096]> var_5311_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583269248))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587463616))))[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4, 1024]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16_palettized, x = var_5311_cast_fp16)[name = string("linear_216_cast_fp16")];
            fp16 var_5316_to_fp16 = const()[name = string("op_5316_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 4, 1024]> var_5317_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5316_to_fp16)[name = string("op_5317_cast_fp16")];
            tensor<fp16, [1, 4, 1024]> input2_2_cast_fp16 = add(x = input1_2_cast_fp16, y = var_5317_cast_fp16)[name = string("input2_2_cast_fp16")];
            tensor<int32, [1]> audio_signal_1_axes_0 = const()[name = string("audio_signal_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587464192)))];
            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587466304)))];
            tensor<fp16, [1, 4, 1024]> audio_signal_1_cast_fp16 = layer_norm(axes = audio_signal_1_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_29_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input2_2_cast_fp16)[name = string("audio_signal_1_cast_fp16")];
            string cast_241_dtype_0 = const()[name = string("cast_241_dtype_0"), val = string("int32")];
            int32 obj1_1_axis_0 = const()[name = string("obj1_1_axis_0"), val = int32(0)];
            tensor<fp16, [24, 1, 56, 1024]> new_cache_last_channel = stack(axis = obj1_1_axis_0, values = (var_484_cast_fp16, var_688_cast_fp16, var_892_cast_fp16, var_1096_cast_fp16, var_1300_cast_fp16, var_1504_cast_fp16, var_1708_cast_fp16, var_1912_cast_fp16, var_2116_cast_fp16, var_2320_cast_fp16, var_2524_cast_fp16, var_2728_cast_fp16, var_2932_cast_fp16, var_3136_cast_fp16, var_3340_cast_fp16, var_3544_cast_fp16, var_3748_cast_fp16, var_3952_cast_fp16, var_4156_cast_fp16, var_4360_cast_fp16, var_4564_cast_fp16, var_4768_cast_fp16, var_4972_cast_fp16, cache_last_channel_cur_1_cast_fp16))[name = string("obj1_1_cast_fp16")];
            int32 obj2_1_axis_0 = const()[name = string("obj2_1_axis_0"), val = int32(0)];
            tensor<fp16, [24, 1, 1024, 8]> new_cache_last_time = stack(axis = obj2_1_axis_0, values = (var_583_cast_fp16, var_787_cast_fp16, var_991_cast_fp16, var_1195_cast_fp16, var_1399_cast_fp16, var_1603_cast_fp16, var_1807_cast_fp16, var_2011_cast_fp16, var_2215_cast_fp16, var_2419_cast_fp16, var_2623_cast_fp16, var_2827_cast_fp16, var_3031_cast_fp16, var_3235_cast_fp16, var_3439_cast_fp16, var_3643_cast_fp16, var_3847_cast_fp16, var_4051_cast_fp16, var_4255_cast_fp16, var_4459_cast_fp16, var_4663_cast_fp16, var_4867_cast_fp16, var_5071_cast_fp16, cache_last_time_cur_1_cast_fp16))[name = string("obj2_1_cast_fp16")];
            tensor<int32, [1]> var_5333 = add(x = cache_last_channel_len, y = max_audio_length_1)[name = string("op_5333")];
            string var_5333_promoted_to_fp16_dtype_0 = const()[name = string("op_5333_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)];
            fp16 var_55_promoted_to_fp16 = const()[name = string("op_55_promoted_to_fp16"), val = fp16(0x1.cp+5)];
            tensor<fp16, [1]> var_5333_to_fp16 = cast(dtype = var_5333_promoted_to_fp16_dtype_0, x = var_5333)[name = string("cast_5")];
            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_55_promoted_to_fp16, x = var_5333_to_fp16)[name = string("clip_1_cast_fp16")];
            string cast_241_promoted_to_fp16_dtype_0 = const()[name = string("cast_241_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_385_to_fp16 = const()[name = string("const_385_to_fp16"), val = fp16(-inf)];
            fp16 var_17_promoted_to_fp16 = const()[name = string("op_17_promoted_to_fp16"), val = fp16(0x1p+2)];
            tensor<int32, [1]> clip_0_cast_fp16_to_int32 = cast(dtype = cast_241_dtype_0, x = clip_0_cast_fp16)[name = string("cast_4")];
            tensor<fp16, [1]> clip_0_cast_fp16_to_int32_to_fp16 = cast(dtype = cast_241_promoted_to_fp16_dtype_0, x = clip_0_cast_fp16_to_int32)[name = string("cast_3")];
            tensor<fp16, [1]> clip_2_cast_fp16 = clip(alpha = const_385_to_fp16, beta = var_17_promoted_to_fp16, x = clip_0_cast_fp16_to_int32_to_fp16)[name = string("clip_2_cast_fp16")];
            tensor<int32, [1]> var_5350_axes_0 = const()[name = string("op_5350_axes_0"), val = tensor<int32, [1]>([1])];
            string language_mask_to_fp16_dtype_0 = const()[name = string("language_mask_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128]> language_mask_to_fp16 = cast(dtype = language_mask_to_fp16_dtype_0, x = language_mask)[name = string("cast_2")];
            tensor<fp16, [1, 1, 128]> var_5350_cast_fp16 = expand_dims(axes = var_5350_axes_0, x = language_mask_to_fp16)[name = string("op_5350_cast_fp16")];
            tensor<int32, [3]> prompt_broadcast_1_reps_0 = const()[name = string("prompt_broadcast_1_reps_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<fp16, [1, 4, 128]> prompt_broadcast_1_cast_fp16 = tile(reps = prompt_broadcast_1_reps_0, x = var_5350_cast_fp16)[name = string("prompt_broadcast_1_cast_fp16")];
            bool input_12_interleave_0 = const()[name = string("input_12_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 1152]> input_12_cast_fp16 = concat(axis = var_18, interleave = input_12_interleave_0, values = (audio_signal_1_cast_fp16, prompt_broadcast_1_cast_fp16))[name = string("input_12_cast_fp16")];
            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587468416))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589827776))))[name = string("prompt_kernel_0_weight_to_fp16_palettized")];
            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589828352)))];
            tensor<fp16, [1, 4, 2048]> linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16_palettized, x = input_12_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<fp16, [1, 4, 2048]> var_5360_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589832512))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591929728))))[name = string("prompt_kernel_2_weight_to_fp16_palettized")];
            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591930304)))];
            tensor<fp16, [1, 4, 1024]> encoded_output = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16_palettized, x = var_5360_cast_fp16)[name = string("linear_218_cast_fp16")];
            string cast_242_dtype_0 = const()[name = string("cast_242_dtype_0"), val = string("int32")];
            string cast_243_dtype_0 = const()[name = string("cast_243_dtype_0"), val = string("int32")];
            tensor<int32, [1]> encoded_length = cast(dtype = cast_242_dtype_0, x = clip_2_cast_fp16)[name = string("cast_0")];
            tensor<int32, [1]> new_cache_last_channel_len = cast(dtype = cast_243_dtype_0, x = clip_1_cast_fp16)[name = string("cast_1")];
        } -> (encoded_output, encoded_length, new_pre_cache, new_cache_last_channel, new_cache_last_time, new_cache_last_channel_len);
}