github MakazhanAlpamys/Soup v0.10.3

latest releases: v0.75.2, v0.75.1, v0.75.0...
6 months ago

What's New

Bug fixes from manual testing report (CPU + trl compatibility):

  • PPO trl >=0.28 compat: PPOTrainer API changed from config= to args= and from manual loop to built-in .train(). Now auto-detects via inspect and supports both APIs
  • PPO use_cpu: Adds use_cpu=True to PPOConfig when no CUDA is available
  • GRPO CPU warning: Warns that GRPO on CPU is experimental (trl generates empty tensors on CPU). Documents as "GPU required"
  • Better error messages: Updated tensor size error to mention GRPO/PPO CPU limitation

Known limitations

  • GRPO on CPU: trl's GRPOTrainer generates empty tensors on CPU (upstream bug). Use a CUDA GPU for GRPO training
  • Table headers on Windows: Rich table borders show ? on Windows cp1252 console (cosmetic, Rich library limitation)

Tests

  • 9 new tests for CPU/API fixes
  • 639 tests total, all passing
  • ruff clean

Install / Upgrade

pip install -U soup-cli

Don't miss a new Soup release

NewReleases is sending notifications on new releases.