code/training/scripts/remote/setup_vllm.sh

#!/usr/bin/env bash
# Install vLLM for serving (run AFTER Phase-0 so a vLLM dep conflict can't
# invalidate the load/train result). vLLM must support the qwen3_5 hybrid arch
# to serve the base model for the secondary-eval baseline.
set -uo pipefail
export PIP_ROOT_USER_ACTION=ignore
export PIP_BREAK_SYSTEM_PACKAGES=1

echo "== installing vllm =="
python -m pip install -U "vllm>=0.23.0" || { echo "VLLM_INSTALL_FAILED" >&2; exit 1; }

echo "== post-install sanity: transformers still loads qwen3.6 config? =="
python - <<'PY'
import transformers
print("transformers", transformers.__version__)
try:
    import vllm; print("vllm", vllm.__version__)
except Exception as e:
    print("vllm import failed:", repr(e))
from transformers import AutoConfig
c = AutoConfig.from_pretrained("Qwen/Qwen3.6-27B", trust_remote_code=True)
print("model_type", getattr(c, "model_type", None), "arch", getattr(c, "architectures", None))
PY
echo "vllm setup done."