How to handle "Could not initialize NNPACK! Reason: Unsupported hardware" warning in PyTorch / Silero VAD on cloud CPU?

I’m running Silero VAD (via PyTorch + torchaudio) on a Linode cloud instance (2 dedicated CPUs, 4 GB RAM). When I process 10-minute audio chunks, I always get repeated warnings like this and it doesn't stop at all (logging it forever):

[W915 13:46:29.003463516 NNPACK.cpp:56] Could not initialize NNPACK! Reason: Unsupported hardware.
[W915 13:46:29.010310416 NNPACK.cpp:56] Could not initialize NNPACK! Reason: Unsupported hardware.

Here is the script that I have in jupyter notebook:

import os
import warnings

import torchaudio

os.environ["USE_NNPACK"] = "0"
warnings.filterwarnings("ignore")
torchaudio.set_audio_backend("sox_io")

import torch
from silero_vad import load_silero_vad, read_audio, get_speech_timestamps

torch.backends.nnpack.enabled = False

model = load_silero_vad()

wav = read_audio('./some-audio.wav')

speech_timestamps = get_speech_timestamps(
  wav,
  model,
  return_seconds=True,
)

None of these helped:

torch.backends.nnpack.enabled = False

os.environ["USE_NNPACK"] = "0"

From what I understand:

NNPACK is an optional acceleration backend for convolutions in PyTorch.
My cloud CPU doesn’t support NNPACK instructions.

What is the correct way to disable or bypass NNPACK in PyTorch so Silero VAD runs normally on CPU without printing these warnings?

Solution

I encountered the same issue. You can solve it by using a CPU-only version of PyTorch:

pip uninstall -y torch torchaudio torchcodec

pip install --pre torch torchaudio torchcodec --index-url https://download.pytorch.org/whl/nightly/cpu

Then you can use torch.hub to load the silero-vad model:

import torch

model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad', model='silero_vad')
(get_speech_timestamps, _, read_audio, _, _) = utils

wav = read_audio('path_to_audio_file')
speech_timestamps = get_speech_timestamps(
  wav,
  model,
  return_seconds=True,  # Return speech timestamps in seconds (default is samples)
)