- detect_model.py picks TP/port/served name/eager/util from config.json
- validate TP against the model config (dense global KV heads, MoE k % 32)
- one apply_patch.py covers both variants, idempotent
- Dockerfile stays generic: no ENV/WORKDIR/EXPOSE/ENTRYPOINT
- run.sh/verify.sh follow the llama4-bi150 layout