[AMD][MI35X] Qwen3.5-fp8 SGLang single-node benchmark

[1am9trash]

Create branch from @yichiche branch.

Successful run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27002889438
Chart: https://inferencex.semianalysis.com/inference?unofficialRun=27002889438

---

Note

Low Risk
Benchmark and CI config only; no application runtime or auth paths. Changes affect published perf numbers and sweep coverage, not production serving.

Overview
Updates the Qwen3.5-397B-A17B-FP8 MI355X SGLang single-node benchmark (base and MTP) to a newer ROCm image (v0.5.12.post1-rocm720-mi35x-20260528) and collapses the fixed-seq-len sweep to a single TP4 / EP1 concurrency range (4–256) for both 1k/1k and 8k/1k, replacing the prior multi-row TP/EP grid.

The launch scripts switch attention from triton to aiter (SGLANG_USE_AITER* env vars), tie --max-running-requests to sweep concurrency, replace --max-prefill-tokens with --chunked-prefill-size 32768, and add multithreaded load plus --page-size 16. perf-changelog.yaml records the image and script changes for both config keys.

^{Reviewed by Cursor Bugbot for commit 930b1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.

[claude]

LGTM — AMD MI355X benchmark image bump and aiter backend retune, no production code paths affected.

Extended reasoning...

Overview

This PR updates the Qwen3.5 FP8 SGLang benchmark recipe for MI355X. It bumps the ROCm image from v0.5.12-rocm720-mi35x-20260517 to v0.5.12.post1-rocm720-mi35x-20260528 in both base and MTP variants of amd-master.yaml, and retunes the two launch scripts: switches attention backend from triton to aiter (with SGLANG_USE_AITER* env vars), drops the per-sweep --cuda-graph-max-bs $CONC and explicit CONTEXT_LENGTH/MAX_PREFILL_TOKENS in favor of --max-running-requests 512, --chunked-prefill-size 32768, --mem-fraction-static 0.9, --page-size 16, and multithreaded model loading.

Security risks

None. This touches only AMD benchmark configuration and SGLang launch flags; no auth, crypto, network exposure, or persisted state is affected. Server is bound to 0.0.0.0 as before — unchanged.

Level of scrutiny

Low. Benchmark recipes for AMD MI355X throughput sweeps — not production-critical code, not user-facing, and not on a code path with broad blast radius. Pattern matches several other recent AMD image-bump PRs (e.g. b28da84 switched the same recipe to aiter previously, 3b685b4 is this image bump). The else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH" branch removal is intentional since CONTEXT_LENGTH is no longer computed.

Other factors

No bugs were flagged by the automated review. Changes are confined to AMD MI355X-specific recipes; other matrix entries are untouched. CI on these benchmark recipes runs on AMD hardware and will surface any launch-flag incompatibilities directly.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27002889438
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27002889438

[chunfangamd]

Attn backend: Triton --> Aiter

Nice work!

[chunfangamd]

@ChangLiu0709 could you please help to update the SGLang cookbook?

[ChangLiu0709]

@ChangLiu0709 could you please help to update the SGLang cookbook?

Sure!

[Oseltamivir]

/reuse-sweep-run

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033542972
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033542972

[Oseltamivir]

lgtm

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27033578807
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27033578807