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

[1am9trash]

Note

Low Risk
Benchmark-only image and launch-flag changes; no application or security-sensitive code paths.

Overview
Updates the Qwen3.5 FP8 MI355X SGLang MTP single-node benchmark to a newer ROCm image and retunes the launch recipe for AITER.

The qwen3.5-fp8-mi355x-sglang-mtp config in .github/configs/amd-master.yaml now points at lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260528 (was ...20260517). benchmarks/single_node/fixed_seq_len/qwen3.5_fp8_mi355x_mtp.sh enables SGLANG_USE_AITER / unified AITER attention, switches --attention-backend from triton to aiter, and adjusts serving limits: fixed --max-running-requests 512, --chunked-prefill-size 32768, --mem-fraction-static 0.9, --page-size 16, multithreaded model load—while dropping the prior ISL/OSL-derived --context-length, --max-prefill-tokens, and concurrency-tied --cuda-graph-max-bs. perf-changelog.yaml documents the config key change.

^{Reviewed by Cursor Bugbot for commit 50efc8f. 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.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 50efc8f. Configure here.}

[cursor]

Missing throughput context length

High Severity

For normal benchmark runs (EVAL_ONLY not true), the server launch no longer receives --context-length. The script drops CONTEXT_LENGTH and the else branch that passed ISL + OSL + 20, so EVAL_CONTEXT_ARGS stays empty while throughput jobs still use fixed ISL/OSL from CI.

 

^{Reviewed by Cursor Bugbot for commit 50efc8f. Configure here.}

[claude]

LGTM — benchmark image bump and AITER launch-flag retune for a single AMD MI355X recipe.

Extended reasoning...

Overview

This PR touches only AMD MI355X benchmark configuration for the qwen3.5-fp8-mi355x-sglang-mtp recipe: a docker image bump in .github/configs/amd-master.yaml (...20260517 → v0.5.12.post1...20260528), a retuned launcher script (qwen3.5_fp8_mi355x_mtp.sh) switching to the AITER attention backend with adjusted serving flags, and a corresponding entry in perf-changelog.yaml.

Security risks

None. No production code paths, auth, crypto, or permissions are touched. Changes are confined to benchmark recipes that run in CI sweeps.

Level of scrutiny

Low. This is a pure benchmark-config change scoped to a single recipe key — well-established pattern for image/flag retuning PRs in this repo. The dropped --context-length (non-EVAL path), --max-prefill-tokens, and --cuda-graph-max-bs are intentional per the PR description and replaced by --chunked-prefill-size and --max-running-requests.

Other factors

• No bugs reported by the bug hunting system.
• Changes are self-contained and follow the same pattern as adjacent recipe entries in the same file.
• The recipe reminder bot comment is the standard reminder, not a blocker.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26993294925
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26993294925