[AMD][MI35X] 0612 DSV4

[1am9trash]

Successful run:
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27406714195
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27406714195

Change:

• Bump image to lmsysorg/sglang-rocm:v0.5.13-rocm720-mi35x-20260612.
• Fix the intermediate_pad setting in the MoE computation in sglang PR#27858. This avoids the unnecessary overhead of computing useless padding.
• Correct the chunk prefill setting size under tp8/dp8 config.

---

Note

Low Risk
Benchmark and launch-script config only; no application auth, data, or core runtime code paths.

Overview
Updates DeepSeek-V4 FP4 on MI355X (SGLang) benchmark wiring for the 0612 ROCm image and aligned server flags.

dsv4-fp4-mi355x-sglang in .github/configs/amd-master.yaml now pins lmsysorg/sglang-rocm:v0.5.13-rocm720-mi35x-20260612 (replacing the prior v0.5.12.post1 build). dsv4_fp4_mi355x_sglang.sh no longer hard-codes --chunked-prefill-size 8192; it uses CHUNKED_PREFILL_SIZE=8192 by default and sets 8192 * TP when DP_ATTENTION is enabled so TP8/DP8 runs use 65536.

perf-changelog.yaml records the image bump, upstream MoE intermediate_pad fix (sglang PR#27858), and the chunked-prefill correction for TP8/DP8.

^{Reviewed by Cursor Bugbot for commit 387b083. 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.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27406714195
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27406714195