Run vLLM on UF HiPerGator through someone else's authenticated SSH session: ControlMaster socket, pipe files over ssh (scp is blocked), sbatch setup then serve
HiPerGator needs password + Duo, which an agent cannot complete. Ask the user to run once: ssh -fNM -S ~/.ssh/cm-hpg -o ControlPersist=24h hpg; afterwards every command uses ssh -S ~/.ssh/cm-hpg -o ControlPath=~/.ssh/cm-hpg -o BatchMode=yes hpg "..." with no credentials. Gotchas: scp over the control socket failed with Connection closed, so send files with ssh hpg "cat > path" < local; sacctmgr shows the QOS limit (here 2 GPUs, 250G mem) and sinfo lists hpg-b200 (8x B200 192GB) and hpg-rtx6000 partitions; login nodes have internet (HF + PyPI 200) and /blue has TBs but home is 40GB, so put HF_HOME, venvs and caches under /blue/<group>/<user>. Pattern that worked: (1) CPU sbatch job creates a venv with module load python/3.11, pip install vllm huggingface_hub hf_transfer, and snapshot_download of Qwen/Qwen3-VL-32B-Instruct (66GB; whole job 27 min; pip resolved vllm 0.33 with torch 2.14+cu130); (2) GPU sbatch job on hpg-b200 runs vllm serve ... --host 0.0.0.0 --api-key $(cat keyfile) --enable-auto-tool-choice --tool-call-parser hermes --limit-mm-per-prompt {"image":6} --max-model-len 32768, key generated on the cluster with chmod 600; (3) reach it from the laptop with ssh -O forward -L 8001:<node>:8000. Watch jobs with a background until-loop on squeue -h -j <id>.