transformers

2f2e711f - [VibeVoice] Add `sample` argument to disable VAE noise in audio features (#50179)

[VibeVoice] Add `sample` argument to disable VAE noise in audio features (#50179) * [VibeVoice-ASR] Add `sample` to disable VAE noise in audio features `get_audio_features` always added `vae_std` noise to the acoustic latents, so outputs could not be made deterministic. Add a `sample` argument (default `True`, unchanged behaviour), plumbed through `forward` so `generate` accepts it too. With `sample=False` the latent mean is used. The common tests now pass `sample=False`, so the nine tests skipped for VAE randomness run again. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> * [VibeVoice] Add `sample` to disable VAE noise in audio features The voice prompt encoder hard-coded `sample=True`. Expose it on `get_audio_features` and both `forward` methods, default `True` (unchanged). With `sample=False` the latents are a view created under `no_grad`, so apply the latent bias and scaling out of place. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> * clearer doc --------- Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com>