| # New-box setup | |
| The vLLM changes now live in the fork **github.com/jarrelscy/vllm-glm52-sm120** | |
| (branch `glm52-sm120`) — everything is committed, there are no patches to apply. | |
| 1. `bash code/run_rtx6000.sh [1m|500k|250k]` — does everything (clones the fork, | |
| builds the venv with `VLLM_USE_PRECOMPILED=1`, installs flashinfer git-main for | |
| the SM120 sparse-MLA path, sets up the JIT toolchain, fetches weights, serves). | |
| - Or use the fork's Docker image (RTX PRO 6000): | |
| `docker build -f Dockerfile.glm52-sm120 -t glm52-sm120 .` then | |
| `docker run --gpus all --ipc=host -p 8001:8001 -v <weights>:/models/1m -e PARALLEL=tp4-1m-mtp glm52-sm120` | |
| (see the Dockerfile header for all `PARALLEL` modes: pp4-1m, tp4-1m-mtp, tp2pp2, …). | |
| 2. Verify with `code/tools/verify_sm120.py` against the golden bundle | |
| (upload it separately or re-capture via `code/tools/capture_golden.py`). | |
| 3. Full runbook: `code/PLAN.md`; history: `code/glm52-RUNLOG.md`. | |