What's New
Bug Fixes
-
PPO resume_from_checkpoint rejected (P1): Fixed
TypeError: PPOTrainer.train() got an unexpected keyword argument 'resume_from_checkpoint'. The trl experimental PPOTrainer.train() has a different signature than standard Trainer.train(). Soup now inspects the signature at runtime and skips the argument when unsupported, printing a warning instead of crashing. -
CPU meta tensor error (P2): Fixed
NotImplementedError: Cannot copy out of meta tensor; no data!. On CPU with no GPU,device_map="auto"can place model weights on a "meta" device, causing crashes when tensors are accessed. All trainers (SFT, DPO, GRPO, PPO, RewardModel) now usedevice_map="cpu"when running on CPU.
Tests
- 12 new tests for both fixes (661 total passing)
Install / Upgrade
pip install --upgrade soup-cli