[GB300][vLLM] Add MiniMax-M2.5 FP4 disagg Dynamo configs#1653
[GB300][vLLM] Add MiniMax-M2.5 FP4 disagg Dynamo configs#1653Ankur-singh wants to merge 2 commits into
Conversation
Adds MiniMax-M2.5 NVFP4 disaggregated multinode vLLM benchmarks on GB300 via Dynamo. Adds model routing + srt-slurm fork path in launch_gb300-nv.sh and launch_gb300-cw.sh, and the 1k1k/8k1k recipe set under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/. Split of #1560 (GB300 half).
|
Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers. If additional help is needed, PR authors can reach out to core maintainers over Slack. |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26856828826 |
|
@Ankur-singh I also have this in progress on #1641 |
|
@jasonlizhengjian Noted. Closing this one! |
2 participants
GB300 half of split #1560 (B300 sibling lives in a separate PR so one CI failure doesn't block the other).
Summary
minimaxm2.5-fp4-gb300-dynamo-vllmto.github/configs/nvidia-master.yaml(1k1k + 8k1k search spaces)benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/minimaxm2.5 + fp4 + dynamo-vllmrouting intorunners/launch_gb300-nv.shandrunners/launch_gb300-cw.shImage:
vllm/vllm-openai:v0.20.1; model:nvidia/MiniMax-M2.5-NVFP4.Note
Low Risk
Benchmark and runner wiring only; no production app or auth changes. Misconfigured paths could fail CI sweeps on GB300 clusters.
Overview
Adds GB300 coverage for MiniMax-M2.5 NVFP4 disaggregated Dynamo + vLLM multinode benchmarks, parallel to the existing GB200 entry.
Registers
minimaxm2.5-fp4-gb300-dynamo-vllminnvidia-master.yamlwith 1k1k and 8k1k fixed-seq-len search spaces (varying prefill/decode workers, TP/EP, andCONFIG_FILErecipe paths). Introduces a new recipe tree underbenchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/(tp4, tp4ep, dep2/dep8, 8k variants) wired tovllm/vllm-openai:v0.20.1, Dynamo 1.2.0 dev wheel, and Nixl KV transfer.Launch scripts on GB300 (
launch_gb300-nv.sh,launch_gb300-cw.sh) now routeminimaxm2.5+fp4+dynamo-vllm: model paths on scratch/vast,srt-slurmclone onmain, recipe copy intorecipes/vllm/minimax-m2.5, andminimax-m2.5-nvfp4insrtslurm.yamlmodel_paths. perf-changelog documents the new config key.Reviewed by Cursor Bugbot for commit dcc0f2b. Bugbot is set up for automated code reviews on this repo. Configure here.