[AMD][MI35X] 0706 DSV4 sglang mtp

[1am9trash]

Successful run:
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28985274062
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28985274062

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28836023139
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28836023139

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28841974603
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28841974603

[seungrokj]

@1am9trash
can you take a look at this ? I canceled the job due to this:
FAIL: gsm8k exact_match,strict-match = 0.8362 (< 0.91 from models.dsv4)
FAIL: gsm8k exact_match,flexible-extract = 0.8415 (< 0.91 from models.dsv4)

https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28841974603/job/85556419336

[1am9trash]

This 2 PR may fix the mtp accuracy issue. Will bump the version after daily image build tomorrow.

• [AMD] Fix DeepSeekV4 server cutlass error sgl-project/sglang#30374
• [AMD] Fix DeepSeek V4 MTP accuracy issue sgl-project/sglang#30333

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28985274062
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28985274062

[1am9trash]

/reuse-sweep-run

[chunfangamd]

As a PR reviewer and CODEOWNER, I have reviewed this and have:

• Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
• Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this. Link: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28985274062
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. Range: 95.53%-96.13%; link: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28985274062
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; FLOPs at lower precisions is fine, given that the config passes private evals. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
• If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
• If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
• Verified that the single-node recipes are similar to the official vLLM recipes and/or theSGLang cookbook:
  • If they are not, I have verified that a PR has been opened in vLLM recipe repo or SGLang repo and linked it below in the additional detail section:
• If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.

Additional detail section:

• insert any additional info here
  SGLang cookbook link: sgl-project/sglang#30651

Signed: @chunfangamd

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — @chunfangamd is a listed owner of configs/amd-master.yaml; remaining paths are catch-all-only, covered by any recognized CODEOWNER.
✅ Check 1 (sweep on in-PR commit): PASS — commit e32361a8 (in this PR) has all single-node */ and eval / check-runs green (executed, not skipped) from run https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28985274062.
✅ Check 2 (evals pass): PASS — gsm8k 0.9553–0.9613 across 3 eval points, all above the dsv4 bar of 0.91 (utils/evals/thresholds.json); run used the same image as this PR's config (lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708).
✅ Check 3 (recipe linked + complete): PASS — SGLang cookbook PR linked (sgl-project/sglang#30651); major args match (DeepSeek-V4-Pro, TP8/DP8 + dp-attention, EAGLE MTP 3-1-4, --attention-backend dsv4, --kv-cache-dtype fp8_e4m3, identical image). Informational only: recipe uses --enable-two-batch-overlap vs this PR's --enable-prefill-delayer, and chunked-prefill-size differs — scheduling/harness tuning, not deployment-defining.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by @1am9trash (COLLABORATOR).
✅ Check 5 (latest checklist): PASS — all current-template items present and checked; the final item is conditional and correctly left unchecked.
✅ Check 6 (upstream image / engine-first): PASS — framework: sglang on MI355X with upstream lmsysorg/sglang-rocm image; engine-first satisfied (this IS the SGLang entry).
✅ Check 7 (no architecture hacks): PASS — no --hf-overrides/override-args; the PR removes an old config.json model_type shim and adds no FLOPs-reducing knobs (fp8 KV cache is precision-only and evals pass).
✅ Check 8 (spec-decode chat template): PASS — client runs with --dsv4, routing prompts through DSv4 chat framing (tokenizer ships no jinja template).

[cquil11]

LGTM