[NVIDIA][SGLang][redo PR] B200 DeepSeek v4 FP4 SGLang: recipe-per-CONC split#1187
[NVIDIA][SGLang][redo PR] B200 DeepSeek v4 FP4 SGLang: recipe-per-CONC split#1187
Conversation
…1158) Mirror of #1185 for the b200 side. Re-applies the b200-specific changes from #1158 on top of the #1186 baseline. - Image pinned to lmsysorg/sglang:deepseek-v4-blackwell@sha256:df18bfc4... - Adds DP_ATTENTION env knob and SGLANG_OPT_* perf env vars - Search space gets conc-start=1 in low-latency rows (was 4) - Recipe-per-CONC dispatch in script: low-latency / balanced / max-throughput selected by DP_ATTENTION + CONC Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
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 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. |
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
cquil11
changed the title
The flag was accidentally dropped during the recipe-per-CONC rewrite. Restoring it to match the baseline methodology (prefix caching disabled) and stay consistent with all other dsv4 sister scripts. Co-authored-by: Cameron Quilici <cquil11@users.noreply.github.com> Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… dispatch The script has two branches (DP_ATTENTION true/false), not three CONC-keyed recipes. Both balanced and max-throughput rows use the same DP-attention + DeepEP flags — only --max-running-requests differs. Updated the nvidia-master.yaml comment block and perf-changelog description to accurately reflect this two-recipe dispatch. Co-authored-by: Cameron Quilici <cquil11@users.noreply.github.com>
2 participants
Summary
Mirror of #1185 for the b200 side. Re-applies the b200-specific changes from #1158 in a single commit on top of the #1186 baseline.
What's included (the b200 part of #1158)
lmsysorg/sglang:deepseek-v4-blackwell@sha256:df18bfc4...nvidia-master.yaml: low-latency rows getconc-start: 1(was 4) for both 1k1k and 8k1kdsv4_fp4_b200.sh:DP_ATTENTIONenv knob incheck_env_varsSGLANG_OPT_*env block (SWA_SPLIT_LEAF_ON_INSERT, USE_JIT_NORM, USE_JIT_INDEXER_METADATA, USE_TOPK_V2, USE_CUSTOM_ALL_REDUCE_V2)DP_ATTENTION+CONCFiles
.github/configs/nvidia-master.yaml—dsv4-fp4-b200-sglangblock updated.benchmarks/single_node/dsv4_fp4_b200.sh— recipe dispatch logic.perf-changelog.yaml— single new entry to trigger the sweep.Test plan
ep=/dpa=suffixes that distinguish the recipes.🤖 Generated with Claude Code