# Deps for the inference step, installed into its venv by command.sh on first run.
# torch==2.12.0: PyPI Linux wheel bundles CUDA 13.0 (GPU); macOS wheel is CPU/MPS.
torch==2.12.0
# >=4.56: run.py uses from_pretrained(dtype=), added in 4.56 (crashes on 4.55.x).
# <5: 4.x apply_chat_template/generate behavior (see run.py generate()).
transformers>=4.56,<5
accelerate
