MakazhanAlpamys/Soup v0.10.1

v0.10.1: Bug Fixes

on GitHub

What's New

Bug fix release based on manual testing report (24 test sections on Windows + CPU).

Fixes

• Windows UnicodeEncodeError — replaced Unicode arrows/dashes with ASCII in all console output (~10 commands affected)
• PPO trl compatibility — uses inspect.signature() to detect parameter names across trl versions (ppo_epochs vs num_ppo_epochs)
• CPU dtype mismatch — new get_compute_dtype() auto-selects float32 on CPU, bfloat16/float16 on CUDA. Fixes reward model and all trainers
• CPU quantization warning — warns when using 4bit/8bit quantization on CPU
• diff command crash — fixed dtype → torch_dtype in model loading
• wandb + trl conflict — pinned wandb<0.18.0 to avoid import errors, added runtime guard

Stats

• 13 new tests (624 total)
• 24 files changed
• All 624 tests pass, ruff clean

Install / Upgrade

pip install --upgrade soup-cli