XNNPACK export of Parakeet TDT model produces different output (mismatch vs NeMo / portable backend)

[sehun1kim-gleo]

🐛 Describe the bug

Hi, I’m encountering an issue when exporting the Parakeet TDT model to ExecuTorch using the XNNPACK backend. The exported model runs, but the transcription output does not match the original NeMo model, while the portable backend works correctly.

I export the model using:

python export_parakeet_tdt.py \
  --audio PATH_TO_LIBRISPEECH/test-clean-processed/1089-134691-0017.wav \
  --backend xnnpack \
  --output-dir OUTPUT_DIR

Here is the full log:

(executorch_clean) [audio@audio-dgx-preproc parakeet]$ python export_parakeet_tdt.py --audio PATH_TO_LIBRISPEECH/test-clean-processed/1089-134691-0017.wav --backend xnnpack --output-dir OUTPUT_DIR
Skipping import of cpp extensions due to incompatible torch version 2.10.0+cu128 for torchao version 0.15.0             Please see https://github.com/pytorch/ao/issues/2919 for more info
I tokenizers:regex.cpp:27] Registering override fallback regex
Extracting tokenizer...
Extracted tokenizer to: OUTPUT_DIR/tokenizer.model
Loading model...
[NeMo W 2026-03-18 07:20:16 megatron_init:62] Megatron num_microbatches_calculator not found, using Apex version.
WARNING:nv_one_logger.api.config:OneLogger: Setting error_handling_strategy to DISABLE_QUIETLY_AND_REPORT_METRIC_ERROR for rank (rank=0) with OneLogger disabled. To override: explicitly set error_handling_strategy parameter.
WARNING:nv_one_logger.training_telemetry.api.training_telemetry_provider:No exporters were provided. This means that no telemetry data will be collected.
[NeMo I 2026-03-18 07:20:41 mixins:184] Tokenizer SentencePieceTokenizer initialized with 8192 tokens
[NeMo W 2026-03-18 07:20:44 modelPT:188] If you intend to do training or fine-tuning, please call the ModelPT.setup_training_data() method and provide a valid configuration file to setup the train data loader.
    Train config : 
    use_lhotse: true
    skip_missing_manifest_entries: true
    input_cfg: null
    tarred_audio_filepaths: null
    manifest_filepath: null
    sample_rate: 16000
    shuffle: true
    num_workers: 2
    pin_memory: true
    max_duration: 10.0
    min_duration: 1.0
    text_field: answer
    batch_duration: null
    max_tps: null
    use_bucketing: true
    bucket_duration_bins: null
    bucket_batch_size: null
    num_buckets: 30
    bucket_buffer_size: 20000
    shuffle_buffer_size: 10000
    
[NeMo W 2026-03-18 07:20:44 modelPT:195] If you intend to do validation, please call the ModelPT.setup_validation_data() or ModelPT.setup_multiple_validation_data() method and provide a valid configuration file to setup the validation data loader(s). 
    Validation config : 
    use_lhotse: true
    manifest_filepath: null
    sample_rate: 16000
    batch_size: 16
    shuffle: false
    max_duration: 40.0
    min_duration: 0.1
    num_workers: 2
    pin_memory: true
    text_field: answer
    
[NeMo I 2026-03-18 07:20:48 rnnt_models:226] Using RNNT Loss : tdt
    Loss tdt_kwargs: {'fastemit_lambda': 0.0, 'clamp': -1.0, 'durations': [0, 1, 2, 3, 4], 'sigma': 0.02, 'omega': 0.1}
[NeMo I 2026-03-18 07:20:48 rnnt_models:226] Using RNNT Loss : tdt
    Loss tdt_kwargs: {'fastemit_lambda': 0.0, 'clamp': -1.0, 'durations': [0, 1, 2, 3, 4], 'sigma': 0.02, 'omega': 0.1}
[NeMo W 2026-03-18 07:20:48 label_looping_base:125] No conditional node support for Cuda.
    Cuda graphs with while loops are disabled, decoding speed will be slower
    Reason: CUDA is not available
[NeMo I 2026-03-18 07:20:48 rnnt_models:226] Using RNNT Loss : tdt
    Loss tdt_kwargs: {'fastemit_lambda': 0.0, 'clamp': -1.0, 'durations': [0, 1, 2, 3, 4], 'sigma': 0.02, 'omega': 0.1}
[NeMo W 2026-03-18 07:20:48 label_looping_base:125] No conditional node support for Cuda.
    Cuda graphs with while loops are disabled, decoding speed will be slower
    Reason: CUDA is not available
[NeMo I 2026-03-18 07:20:49 save_restore_connector:285] Model EncDecRNNTBPEModel was successfully restored from /home/audio/.cache/huggingface/hub/models--nvidia--parakeet-tdt-0.6b-v3/snapshots/6d590f77001d318fb17a0b5bf7ee329a91b52598/parakeet-tdt-0.6b-v3.nemo.

Exporting components...
[NeMo W 2026-03-18 07:20:57 nemo_logging:364] /mnt/audio/users/sehun-kim/miniconda3/envs/executorch_clean/lib/python3.11/contextlib.py:144: UserWarning: The tensor attributes self.decoder.prediction.dec_rnn.lstm._flat_weights[0], self.decoder.prediction.dec_rnn.lstm._flat_weights[1], self.decoder.prediction.dec_rnn.lstm._flat_weights[2], self.decoder.prediction.dec_rnn.lstm._flat_weights[3], self.decoder.prediction.dec_rnn.lstm._flat_weights[4], self.decoder.prediction.dec_rnn.lstm._flat_weights[5], self.decoder.prediction.dec_rnn.lstm._flat_weights[6], self.decoder.prediction.dec_rnn.lstm._flat_weights[7] were assigned during export. Such attributes must be registered as buffers using the `register_buffer` API (https://pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module.register_buffer).
      next(self.gen)
    

Lowering to ExecuTorch with XNNPACK...

Saving ExecuTorch program to: OUTPUT_DIR/model.pte
Saved 2432.2 MB

============================================================
Testing transcription...
============================================================

[Eager PyTorch]
  Result: The Europe they had come from lay out there beyond the Irish sea, Europe of strange tongues and valleyd and wood begirt and citadelled and of entrenched and marshalled races.

[ExecuTorch Runtime]
[program.cpp:153] InternalConsistency verification requested but not available
[cpuinfo_utils.cpp:71] Reading file /sys/devices/soc0/image_version
[cpuinfo_utils.cpp:87] Failed to open midr file /sys/devices/soc0/image_version
[cpuinfo_utils.cpp:100] Reading file /sys/devices/system/cpu/cpu0/regs/identification/midr_el1
[cpuinfo_utils.cpp:109] Failed to open midr file /sys/devices/system/cpu/cpu0/regs/identification/midr_el1
[cpuinfo_utils.cpp:125] CPU info and manual query on # of cpus dont match.
  Result: Europe of strange tongues and valley and wood figure and citadels and of entrenchment and martial races.

✗ Transcriptions differ!
  Eager: The Europe they had come from lay out there beyond the Irish sea, Europe of strange tongues and valleyd and wood begirt and citadelled and of entrenched and marshalled races.
  ET:    Europe of strange tongues and valley and wood figure and citadels and of entrenchment and martial races.

Done!

I get correct output when I export the model with protable backend.
Could this be related to unsupported ops or silent fallback behavior in XNNPACK?
Thanks in advance for your help!

Versions

PyTorch version: 2.10.0+cu128
Is debug build: False
CUDA used to build PyTorch: 12.8
ROCM used to build PyTorch: N/A

OS: Amazon Linux 2023.8.20250908 (x86_64)
GCC version: (GCC) 11.5.0 20240719 (Red Hat 11.5.0-5)
Clang version: Could not collect
CMake version: version 3.22.2
Libc version: glibc-2.34

Python version: 3.11.15 (main, Mar 11 2026, 17:20:07) [GCC 14.3.0] (64-bit runtime)
Python platform: Linux-6.1.148-173.267.amzn2023.x86_64-x86_64-with-glibc2.34
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8488C
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 16
Socket(s): 1
Stepping: 8
BogoMIPS: 4800.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd ida arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid cldemote movdiri movdir64b md_clear serialize amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 768 KiB (16 instances)
L1i cache: 512 KiB (16 instances)
L2 cache: 32 MiB (16 instances)
L3 cache: 105 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
Vulnerability Gather data sampling: Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsa: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] executorch==1.1.0
[pip3] numpy==2.4.3
[pip3] nv-one-logger-pytorch-lightning-integration==2.3.1
[pip3] nvidia-cublas-cu12==12.8.4.1
[pip3] nvidia-cuda-cupti-cu12==12.8.90
[pip3] nvidia-cuda-nvrtc-cu12==12.8.93
[pip3] nvidia-cuda-runtime-cu12==12.8.90
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cufft-cu12==11.3.3.83
[pip3] nvidia-curand-cu12==10.3.9.90
[pip3] nvidia-cusolver-cu12==11.7.3.90
[pip3] nvidia-cusparse-cu12==12.5.8.93
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] onnx==1.20.1
[pip3] pytorch-lightning==2.6.1
[pip3] pytorch-metric-learning==2.9.0
[pip3] pytorch_tokenizers==1.1.0
[pip3] torch==2.10.0
[pip3] torch-audiomentations==0.12.0
[pip3] torch_pitch_shift==1.2.5
[pip3] torchao==0.15.0
[pip3] torchaudio==2.10.0
[pip3] torchcodec==0.10.0
[pip3] torchmetrics==1.9.0
[pip3] triton==3.6.0
[conda] executorch 1.1.0 pypi_0 pypi
[conda] numpy 2.4.3 pypi_0 pypi
[conda] nv-one-logger-pytorch-lightning-integration 2.3.1 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.8.4.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.8.90 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.8.93 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.8.90 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.10.2.21 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.3.3.83 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.9.90 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.7.3.90 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.5.8.93 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.7.1 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.27.5 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.8.93 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.8.90 pypi_0 pypi
[conda] pytorch-lightning 2.6.1 pypi_0 pypi
[conda] pytorch-metric-learning 2.9.0 pypi_0 pypi
[conda] pytorch-tokenizers 1.1.0 pypi_0 pypi
[conda] torch 2.10.0 pypi_0 pypi
[conda] torch-audiomentations 0.12.0 pypi_0 pypi
[conda] torch-pitch-shift 1.2.5 pypi_0 pypi
[conda] torchao 0.15.0 pypi_0 pypi
[conda] torchaudio 2.10.0 pypi_0 pypi
[conda] torchcodec 0.10.0 pypi_0 pypi
[conda] torchmetrics 1.9.0 pypi_0 pypi
[conda] triton 3.6.0 pypi_0 pypi