[NV] Update Qwen3.5 FP4 B200 SGLang

[Ankur-singh]

Summary

Update Qwen3.5 FP4 B200 SGLang benchmark configuration and launch script for improved throughput sweeps and tuned server parameters.

Config changes (.github/configs/nvidia-master.yaml)

• Image update: lmsysorg/sglang:v0.5.10.post1-cu130 → lmsysorg/sglang:nightly-dev-20260422-de962f32
• Search space: Add tp: 2, ep: 1, conc: 4-128 sweep for both 1k1k and 8k1k sequence lengths
• Baseline: Keep tp: 4, ep: 1, conc: 4 as low-latency baseline

Script changes (benchmarks/single_node/qwen3.5_fp4_b200.sh)

• Remove --max-running-requests $CONC (let SGLang auto-manage)
• Reduce --max-prefill-tokens and --chunked-prefill-size from 81920 → 16384
• Dynamic --scheduler-recv-interval: 30 when CONC > 4, 10 otherwise
• Existing flags retained: --enable-symm-mem, --tokenizer-path, --mem-fraction-static 0.8, --stream-interval 50

Changelog (perf-changelog.yaml)

• Updated entry for qwen3.5-fp4-b200-sglang reflecting all above changes

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

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

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sgl-cookbook

[claude]

This PR is marked [WIP] — deferring for human review once it's ready.

Extended reasoning...

Overview

The PR modifies two files: the nvidia-master.yaml CI config (adding a tp=2 search space entry and narrowing tp=4 to conc=4 only) and the qwen3.5_fp4_b200.sh benchmark launch script (removing deprecated env exports and conditional TP8 logic, updating server flags including --expert-parallel-size, --enable-symm-mem, --tokenizer-path, and tuning memory/prefill parameters).

Security Risks

No security risks. These are benchmark configuration and shell script changes with no auth, crypto, or data-access implications.

Level of Scrutiny

Low-risk changes — benchmark scripts and CI sweep configs are not production-critical paths. The changes are consistent in intent and follow established patterns in the repo.

Other Factors

The PR title includes [WIP], indicating the author considers it incomplete. No bugs were found. A human should confirm the PR is finished before merging.

[gracehonv]

Please change:

• { tp: 2, ep: 2, conc-start: 4, conc-end: 128 }
  to:
• { tp: 2, ep: 1, conc-start: 4, conc-end: 128 }

[gracehonv]

{ tp: 2, ep: 2, conc-start: 4, conc-end: 128 }
to:
{ tp: 2, ep: 1, conc-start: 4, conc-end: 128 }

[hshrivastava-droid]

@claude - add PR description

[Klaud-Cold]

Claude finished @hshrivastava-droid's task in 1m 3s —— View job

---

Adding PR Description

• Gather context on PR changes
• Write and update PR description

PR description has been added with a summary of all changes across the three modified files:

• Config: image update + TP2 sweep search space
• Script: tuned server launch flags (prefill, scheduler-recv-interval, removed max-running-requests)
• Changelog: updated perf-changelog.yaml entry

[gracehonv]

LGTM

[functionstackx]

lgtm. ty for the recipe sgl-project/sgl-cookbook#264

[functionstackx]

@Ankur-singh one small nit, dont remove whitespace plz or it will break

[jgangani]

LGTM

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=25232812945
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=25232812945