What's New
PPO trl >=0.28 API compatibility fix:
- PPO
args=vsconfig=: trl 0.28+ renamed PPOTrainer'sconfig=toargs=andtokenizer=toprocessing_class=. Now auto-detects viainspect.signature - PPO built-in
.train(): trl 0.28+ PPOTrainer inherits from OnlineDPOTrainer with a built-in.train()method. Now detects and uses it instead of the manual generate+step loop - PPO
use_cpu: Addsuse_cpu=Trueto PPOConfig on CPU-only systems - GRPO CPU: Warning that GRPO on CPU is experimental (trl upstream bug). Documented as "GPU required"
- Better error messages: Tensor size errors now mention GRPO/PPO CPU limitation
Known limitations
- GRPO on CPU: trl's GRPOTrainer generates empty tensors on CPU (upstream bug). Use a CUDA GPU
- Table borders on Windows: Rich uses Unicode box-drawing chars that show as
?on cp1252 (cosmetic)
Tests
- 639 tests, all passing, ruff clean
Install / Upgrade
pip install -U soup-cli