Skip to content

CPU can incur a slow path on non-contiguous magnitudes #47351

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
vasqu merged 7 commits into from
Jul 22, 2026
+44 −1
Merged

CPU can incur a slow path on non-contiguous magnitudes #47351

Changes from 1 commit
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
0dd8a71
Make Whisper mel magnitudes contiguous to avoid slow strided matmul
vbayanag Jul 15, 2026
346cc0c
Merge branch 'main' into whisper-mel-contiguous-cpu-speedup
vbayanag Jul 16, 2026
bf2f2a8
Added test case to assert and benchmark
vbayanag Jul 17, 2026
e6debd2
enabled warmup in test case
vbayanag Jul 17, 2026
dce485b
Remove median timing function from tests
vbayanag Jul 21, 2026
bcba305
Apply make style (ruff)
vbayanag Jul 21, 2026
d756166
Reference PR #47351 in test docstring; drop stale -s note
vbayanag Jul 22, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion src/transformers/models/whisper/feature_extraction_whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -147,7 +147,11 @@ def _torch_extract_fbank_features(self, waveform: np.ndarray, device: str = "cpu
waveform += self.dither * torch.randn(waveform.shape, dtype=waveform.dtype, device=waveform.device)

stft = torch.stft(waveform, self.n_fft, self.hop_length, window=window, return_complex=True)
magnitudes = stft[..., :-1].abs() ** 2
# `stft[..., :-1]` is a non-contiguous view; on some CPU backends the
# downstream `mel_filters.T @ magnitudes` matmul hits a slow strided
# path (observed ~8x slower). Forcing contiguity here is cheap and
# keeps the matmul on the fast path.
magnitudes = (stft[..., :-1].abs() ** 2).contiguous()

mel_filters = torch.from_numpy(self.mel_filters).to(device, torch.float32)
mel_spec = mel_filters.T @ magnitudes
Expand Down
Loading