Skip to content

Conversation

sskalnik
Copy link

@sskalnik sskalnik commented Oct 8, 2025

  • Add experimental support for google/t5gemma-b-b-ul2.
  • Add experimental support for pytorch_optimizer.lr_scheduler.chebyshev.
  • Add support for AdamW8bit from bitsandbytes. Improve custom metadata module logic. Fix issue with FutureWarning about torch.cuda.amp.autocast() vs torch.amp.autocast("cuda").
  • Create TRAINING_FOR_NOOBS.md
  • Use dill for PreEncodedDataset because pickling fails on Windows 11.
  • Add AIF and AIFF support. Use dill because pickling fails on Windows 11.
  • Upgrade pytorch_lightning 2.1.0 to 2.4.0
  • Add example config files and training scripts.

sskalnik and others added 15 commits September 13, 2025 19:57
…scripts. Update .gitignore to ignore raw audio file dataset folder and outputs folder.
…module logic. Fix issue with FutureWarning about torch.cuda.amp.autocast() vs torch.amp.autocast("cuda").
…e_size": 354304 instead of 352800. The former is divisible by 2048, slighty over 8 seconds. The latter is exactly 8 seconds divided by 44100, not divisible by 2048.
…e StatefulDataLoader from torchdata so we can resume from mid-epoch checkpoints.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant