dsr1-fp4-b200-dynamo-sglang-mtp: 8k1k 6-variant MTP sweep on local split recipes#1688
dsr1-fp4-b200-dynamo-sglang-mtp: 8k1k 6-variant MTP sweep on local split recipes#1688Ankur-singh wants to merge 5 commits into
Conversation
…ocal split recipes
|
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=27162981754 |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
…tinode runner Only the 8k1k scenario is updated (6-variant local split recipes). The 1k1k scenario and the b200-multinode runner are unchanged from main; the image bump to v0.5.12.post1 is shared (1k1k follows via the dynamo-sglang container alias).
…cipes Flatten the srt-slurm b200-fp4 1k1k recipe (base + zip_override_mtp_*[i]) into 4 standalone per-topology recipes under recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/, matching the 8k1k local layout, and point the config at them instead of srt-slurm. Behavior is unchanged (faithful flatten; dynamo-sglang container alias preserved).
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 79a9f56. Configure here.
| cd "$SRT_REPO_DIR" || exit 1 | ||
| git checkout main | ||
| mkdir -p recipes/sglang/dsr1/b200-fp4 | ||
| cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4" recipes/sglang/dsr1/b200-fp4 |
There was a problem hiding this comment.
Launch branch breaks non-MTP
High Severity
The new dynamo-sglang + dsr1 + fp4 branch applies to every DeepSeek-R1 FP4 disagg run, not only the MTP sweep. It checks out srt-slurm main and copies only recipes/sglang/dsr1/b200-fp4, while dsr1-fp4-b200-dynamo-sglang still points CONFIG_FILE at recipes/b200-fp4/1k1k.yaml overrides that lived on sa-submission-q2-2026.
Reviewed by Cursor Bugbot for commit 79a9f56. Configure here.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27163702669 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27165708753 |
…hash Add dynamo.hash=5b4bc1dd70965017a737c71b19db5a0aeaa88727 + install: true to the four 1k1k MTP recipes so the 1k1k and 8k1k scenarios build dynamo from an identical revision; the 1k1k recipes keep the dynamo-sglang container alias. Update the perf-changelog entry to note the pin.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27229634079 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27229634079 |
2 participants
Restructure the DeepSeek-R1 FP4 B200 dynamo-sglang MTP disagg sweep to an 8k1k-only, 6-variant configuration backed by local split recipes (one flat recipe YAML per topology).
lmsysorg/sglang:v0.5.12.post1dsr1/fp4recipe-copy path tolaunch_b200-dgxc.shNote
Medium Risk
Large benchmark/topology and container-version changes affect multi-node GPU job scheduling and runtime behavior, though scope is limited to the dsr1-fp4-b200-dynamo-sglang-mtp config and launch plumbing.
Overview
Moves the dsr1-fp4-b200-dynamo-sglang-mtp sweep off srt-slurm zip overrides onto local flat recipe YAMLs under
benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/, and pointsnvidia-master.yamlat those paths for both 1k1k (four MTP variants, same topologies as before) and 8k1k.For 8k1k, the search space is reorganized into six explicit disagg topologies: three low-latency (1p5d / 1p3d / 1p1d with TP4 prefill, TP8 decode,
ep: 1) and three MTP2 throughput Pareto points (2p1d / 3p1d / 5p1d with DEP4 prefill / DEP8 decode and fixed high concurrencies). Recipes pin Dynamo todynamo.hash=5b4bc1dd…withinstall: truefor 1k1k as well as 8k1k.Bumps the config image to
lmsysorg/sglang:v0.5.12.post1, documents the change inperf-changelog.yaml, and adds alaunch_b200-dgxc.shbranch that clones srt-slurm@main and copies the localdsr1/b200-fp4recipes into the checkout before runs.Reviewed by Cursor Bugbot for commit d47c291. Bugbot is set up for automated code reviews on this repo. Configure here.