docs: add Modal GPU serving guide

[felipeocampoos]

Summary

• How we used Modal as a drop-in replacement for MIT SLURM GPU partitions during the BODHI-Medcalc rerun campaign, when MIT's GPU QOS was maxed out
• One-class-per-model vLLM serving pattern, deployable in a few minutes
• The seven gotchas that actually cost time: silent GPU-quota exhaustion (max_containers), TP=2+AWQ hangs, incomplete client venvs, the checkpoint-hash relaunch trap, per-model SLURM time limits, HF Hub rate-limiting, NFS cache races

Why

Sebastian asked for a shareable writeup so others on the team can use Modal when the MIT GPU queue is full, without rediscovering the same gotchas.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[maximinl]

Solid, useful writeup. The one-class-per-model pattern, the max_containers quota gotcha, and the checkpoint-hash relaunch trap are exactly the kind of institutional knowledge that otherwise lives in Slack threads.

One soft note for benchmaxxing readers: this repo’s OpenAI-compatible dispatch (experiments/_lane.py and siblings) reads BENCHMAXXING_LOCAL_BASE_URL, not OPENAI_COMPATIBLE_BASE_URL. If the BODHI-Medcalc client really used the latter, worth a one-line “your harness may name it differently; here it’s …” so people don’t export the wrong var and wonder why nothing moves. Not a blocker for landing the doc as campaign notes.

Approve.