Add Qwen3.5 FP8 B200 SGLang MTP config

[ankursingh-nv]

Summary

Adds a new benchmark configuration for Qwen3.5-397B-A17B FP8 on B200 using SGLang with MTP (Multi-Token Prediction) via EAGLE speculative decoding.

Changes

New benchmark script: benchmarks/single_node/qwen3.5_fp8_b200_mtp.sh

• SGLang launch with FP8 quantization and FP8 E4M3 KV cache
• EAGLE speculative decoding config: num-steps=3, draft-tokens=4, topk=1
• FlashInfer + TRT-LLM backends (fp8-gemm-backend=flashinfer_trtllm, attention-backend=trtllm_mha, moe-runner-backend=flashinfer_trtllm)
• FlashInfer allreduce fusion enabled
• Adaptive scheduler recv interval (10 for low concurrency, 30 for CONC >= 16)
• Radix cache disabled, context length dynamically set from ISL + OSL

Config entry: qwen3.5-fp8-b200-sglang-mtp in nvidia-master.yaml

• Image: lmsysorg/sglang:v0.5.9-cu130
• Model: Qwen/Qwen3.5-397B-A17B-FP8
• Single-node, TP=4, EP=1
• Concurrency sweep: 4–256 across all three sequence-length configs (1k/1k, 1k/8k, 8k/1k)

Changelog entry in perf-changelog.yaml

• Documents the new config addition with spec-decoding details

Test Plan

• Run e2e benchmarks for qwen3.5-fp8-b200-sglang-mtp to validate server startup and benchmark completion
• Verify MTP speculative decoding produces correct results across sequence lengths
• Compare throughput/latency against the non-MTP qwen3.5-fp8-b200-sglang baseline

[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

[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

[functionstackx]

.

[cquil11]

@claude add PR description

[Klaud-Cold]

Claude finished @cquil11's task in 49s —— View job

---

Adding PR Description

• Read changed files to understand the PR
• Write and update PR description

PR description has been added. It covers the new benchmark script, the nvidia-master.yaml config entry, and the changelog update.

[ankursingh-nv]

@jgangani @functionstackx can you guys please review and approve the PR?

[jgangani]

@jgangani @functionstackx can you guys please review and approve the PR?

@Ankur-singh Looks good to me. Would be good to revisit later EAGLE draft tokens < 4 for higher concurrencies.