docker/requirements_docker.txt

# Versions match Colab GPU runtime (verified 2026-05). torch + torchvision
# are NOT listed here — they come from the pytorch/pytorch base image with
# the right CUDA build baked in.
#
# Pins are loose (>=) on purpose: pip resolves the latest mutually-compatible
# set. If you want to lock exact versions for reproducibility, generate a
# constraints file from a working build with:
#     pip freeze > requirements_lock.txt

# ── Core HF stack ─────────────────────────────────────────────────────────────
transformers>=4.46,<4.50
peft>=0.13,<0.15
accelerate>=1.0
bitsandbytes>=0.45
huggingface_hub>=0.27,<1.0   # transformers 4.49 declares hf_hub<1.0 — must match
httpx>=0.27                   # utils/_httpx_compat.py handles 0.28+ kwarg change

# ── Vision encoder ────────────────────────────────────────────────────────────
# rad_dino loads via transformers AutoModel — no extra dep needed.
# hi-ml-multimodal (BioViL-T) intentionally omitted; model/rad_dino.py wraps
# its import in try/except and falls back to timm/transformers cleanly.
timm>=1.0
Pillow>=10.0

# ── Config / data ─────────────────────────────────────────────────────────────
omegaconf==2.3.0
sentencepiece>=0.2
protobuf>=5.0
numpy>=1.26,<3        # cap at 2.x for now; some libs still don't support numpy 2
pandas>=2.2

# ── Eval metrics ──────────────────────────────────────────────────────────────
nltk>=3.9
rouge-score>=0.1.2
bert-score>=0.3.12
scikit-learn>=1.5

# ── Training / experiment tracking ────────────────────────────────────────────
wandb>=0.18
tqdm>=4.66

# ── Optional: LLM-as-judge for VQA ────────────────────────────────────────────
openai>=1.40