[Klaud Cold] Add dsr1-fp4-b200-sglang-mtp single-node MTP recipe

[functionstackx]

Summary

Add the single-node MTP/EAGLE speculative-decoding sibling for the existing dsr1-fp4-b200-sglang recipe.

Field	Value
Recipe key	dsr1-fp4-b200-sglang-mtp
Model	nvidia/DeepSeek-R1-0528-FP4-V2 (same as the off sibling)
Image	lmsysorg/sglang:v0.5.12-cu130 (same as the off sibling)
Search space	tp=8 ep=1 conc 4..512 spec-decoding=mtp on 1k1k + 8k1k (matches dsr1-fp8-b200-sglang-mtp)

Launch script

benchmarks/single_node/dsr1_fp4_b200_mtp.sh clones the off variant (dsr1_fp4_b200.sh) and overlays MTP bits from the production B200 sglang MTP template (dsr1_fp8_b200_mtp.sh):

• TP=8 enforcement check
• --cuda-graph-max-bs 512 / --max-running-requests 512 (up from 256 in the off variant)
• --speculative-algorithm EAGLE --speculative-num-steps 2 --speculative-num-draft-tokens 3 --speculative-eagle-topk 1
• Env: SGLANG_ENABLE_SPEC_V2=1
• --use-chat-template on the bench client

Keeps fp4-specific bits intact: --quantization modelopt_fp4, --moe-runner-backend flashinfer_trtllm, --attention-backend trtllm_mla.

Test plan

• YAML loads; bash -n syntax passes on the new launch script.
• full-sweep-enabled sweep finishes green on b200 for tp=8 ep=1 conc 4..512 spec-decoding=mtp / 1k1k + 8k1k.

[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://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

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

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26135019453
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26135019453

[claude]

LGTM — straightforward benchmark recipe addition that cleanly combines the existing fp4 off variant with the fp8 MTP template.

Extended reasoning...

Overview

PR adds a new single-node benchmark recipe dsr1-fp4-b200-sglang-mtp consisting of: (1) a new YAML entry in .github/configs/nvidia-master.yaml, (2) a new launch script benchmarks/single_node/dsr1_fp4_b200_mtp.sh that combines the fp4 off-variant with MTP/EAGLE speculative-decoding flags from the existing fp8 MTP template, and (3) a corresponding changelog entry. Scope is purely additive — no existing recipes are modified.

Security risks

None. This is a benchmark configuration/launch script that runs in a sandboxed CI environment against a public model image.

Level of scrutiny

Low. This is a mechanical, pattern-following recipe addition that closely mirrors two existing scripts (dsr1_fp4_b200.sh for fp4 specifics and dsr1_fp8_b200_mtp.sh for MTP overlay). The differences are intentional and match the description: TP=8 check, bumped cuda-graph-max-bs/max-running-requests to 512, MTP/EAGLE flags, SGLANG_ENABLE_SPEC_V2=1, and --use-chat-template on the bench client. The fp4-specific bits (modelopt_fp4, flashinfer_trtllm, trtllm_mla) are preserved.

Other factors

The bug hunting system found no issues. The PR is labeled full-sweep-enabled so the actual sweep will validate correctness empirically. Minor nice-to-have: the new script properly uses export SGLANG_ENABLE_SPEC_V2=1 (the fp8 template omits export), which is actually a slight improvement.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26135021096
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26135021096

[functionstackx]

/reuse-sweep-run

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26144409401
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26144409401