MakazhanAlpamys/Soup v0.10.6

on GitHub

What's New

Bug Fixes

• PPO trl >=0.28 experimental API (P1): Fixed TypeError: PPOTrainer.__init__() missing 4 required positional arguments. trl >=0.28 moved PPOTrainer to trl.experimental.ppo with a new API requiring ref_model, reward_model, train_dataset, and value_model. Soup now auto-imports from trl.experimental.ppo with fallback, and creates reward/value models when needed.

• GRPO CPU empty generation (P2): Fixed RuntimeError: The expanded size of the tensor (2) must match the existing size (0). trl's GRPOTrainer generates empty completion tensors on CPU, causing tensor size mismatch. Soup now passes generation_kwargs={"min_new_tokens": 1} on CPU to ensure at least one token is generated.

Tests

• 6 new tests for both fixes (649 total passing)

Install / Upgrade

pip install --upgrade soup-cli