{{ message }}
Repository navigation
Merge #16230 to main - #16382
Open
blisc wants to merge 5 commits into
Open
Merge #16230 to main#16382blisc wants to merge 5 commits into
blisc wants to merge 5 commits into
Conversation
* dont save audios on disk Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Train predicted phonemes with GRPO Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Prevent NaNs in phoneme GRPO masking Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Decouple phoneme rollout sampling Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Match audio rollout likelihood in GRPO Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Disable top-k for configured GRPO rollouts Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Simplify phoneme sampling for GRPO Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> * Revert "Simplify phoneme sampling for GRPO" This reverts commit a40dc17. * Remove local GRPO unit tests Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> --------- Signed-off-by: Shehzeen Hussain <shehzeenh@nvidia.com> (cherry picked from commit d0b2f63) Signed-off-by: Jason <jasoli@nvidia.com>
Signed-off-by: Jason <jasoli@nvidia.com>
1 of 17 tasks
Signed-off-by: Jason <jasoli@nvidia.com>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
| def _remove_directory(directory: str) -> None: | ||
| try: | ||
| shutil.rmtree(directory) | ||
| except FileNotFoundError: |
…ures Signed-off-by: Jason <jasoli@nvidia.com>
…outputs Signed-off-by: Jason <jasoli@nvidia.com>
blisc
commented
Oct 19, 2026
Comment on lines
+470
to
+471
| if not self.cfg.get('save_reward_audios', True): | ||
| raise RuntimeError('Reward audio saving is disabled.') |
Collaborator
Author
There was a problem hiding this comment.
@shehzeen: Assuming this works, should we strip out all saving to disk functionality from this class?
shehzeen
reviewed
Oct 20, 2026
Collaborator
There was a problem hiding this comment.
Could we preserve the previous UTMOS preprocessing behavior here? process_directory() called upstream model.predict(), which defaults to remove_silent_section=True. The new direct get_dataset() path bypasses that override and appears to leave silence removal disabled, changing the MOS reward for generations with leading/trailing pauses.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Bring predicted-phoneme online preference optimization from #16230 to main. Reuse the sampled phoneme trajectory during teacher forcing and optimize it jointly with audio using a configurable loss weight.
Generated and reference reward WAVs previously remained in temporary storage after scoring. With
save_reward_audios=False, each reward call now owns a unique temporary directory and removes its files when it returns or raises. Saved-audio mode retains its existing behavior.Collection: TTS
Changelog
Usage
Set
model.save_reward_audios=falsefor per-call cleanup. For predicted-phoneme PO, configure:Additional information
Source PR: #16230, plus the requested per-iteration artifact cleanup. The objective, CFG, codec batching, curriculum, and reward-ASR changes introduced by #16301 are kept in the subsequent
tts-reward-asrPR.