[WIP] Update Qwen3.5 FP8 B200 SGLang

[Ankur-singh]

Summary

• Search-space: tp8 conc=4, tp4 conc=4-256 (both seq-lens)
• Updated server launch args to match latest recipe

[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

[Ankur-singh]

@claude update perf-changelog.yaml file

[Klaud-Cold]

Claude finished @Ankur-singh's task in 56s —— View job

---

Updated perf-changelog.yaml

• Gather context on PR changes
• Read current perf-changelog.yaml
• Add changelog entry
• Commit and push

Added a perf-changelog.yaml entry for qwen3.5-fp8-b200-sglang covering:

• Search-space update (tp8 conc=4, tp4 conc=4-256; ep removed)
• New server args (--enable-symm-mem, --disable-radix-cache, --tokenizer-path)
• Tuning changes (prefill 32768→16384, mem-fraction 0.82→0.8, stream-interval 30→50)
• Removed env vars and conditional --enable-flashinfer-allreduce-fusion

Commit: 2b15d3d
|

[claude]

This PR is still marked [WIP] and also removes the ep field from the search-space YAML while the benchmark script still requires EP_SIZE as a mandatory env var — worth a human look before merging.

Extended reasoning...

Overview

Two files changed: the nvidia-master.yaml CI config (search-space parameters for qwen3.5-fp8-b200-sglang) and benchmarks/single_node/qwen3.5_fp8_b200.sh (server launch args). The changes simplify the server invocation and narrow the tp8 concurrency sweep to a single point (4→4) while widening tp4 (4–256).

Security Risks

No security-sensitive code is touched. No auth, crypto, or permissions changes.

Level of Scrutiny

Moderate. This is benchmark/CI config code, not production logic, but changes to the search-space and server launch args affect the validity of benchmark results and CI correctness. The [WIP] title explicitly signals the author considers it unfinished.

Other Factors

The ep key is removed from both search-space entries in the YAML, yet the benchmark script still lists EP_SIZE in its check_env_vars call (mandatory) and passes it as --expert-parallel-size=$EP_SIZE. If EP_SIZE is not injected by the CI harness through another path, jobs will fail at startup. This potential misalignment, combined with the [WIP] designation, warrants a human review before merging.