[AMD][MI35X]Update qwen3.5 perf

[zhentaocc]

Co-authored-by: @chunfangamd @1am9trash

• update image, include pr changes:
  • [AMD] Add fused GemmaRMSNorm forward_hip to use aiter/vllm kernels for qwen3.5 sgl-project/sglang#21188
  • [AMD]Integrate aiter's fused_topk for softmax scoring in topk function sgl-project/sglang#21421
  • [AMD] Enable share expert fusion with router experts for Qwen3.5 BF16 & FP8 sgl-project/sglang#20736
• update command args

[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://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your 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/sgl-cookbook

[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://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your 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/sgl-cookbook

[functionstackx]

@claude can u add these links to changelog yaml

pdate image, include pr changes:
sgl-project/sglang#21188
sgl-project/sglang#21421
sgl-project/sglang#20736

[Klaud-Cold]

Claude finished @functionstackx's task in 50s —— View job

---

Adding SGLang PR links to perf-changelog.yaml

• Read current perf-changelog.yaml and PR diff
• Add SGLang PR links to changelog entries
• Commit and push changes

Added the three upstream SGLang PR links to the existing qwen3.5-fp8-mi355x-sglang / qwen3.5-bf16-mi355x-sglang changelog entry:

• [AMD] Add fused GemmaRMSNorm forward_hip to use aiter/vllm kernels for qwen3.5 sgl-project/sglang#21188
• [AMD]Integrate aiter's fused_topk for softmax scoring in topk function sgl-project/sglang#21421
• [AMD] Enable share expert fusion with router experts for Qwen3.5 BF16 & FP8 sgl-project/sglang#20736

Commit: 8c51a630
|

[claude]

---

Additional findings (outside current diff — PR may have been updated during review):

• 🟡 perf-changelog.yaml:1362-1369 — The new perf-changelog entry has three copy-paste errors: (1) qwen3.5-fp8-mi355x-sglang is listed twice in config-keys (positions 0 and 2) while qwen3.5-bf16-mi355x-sglang appears only once — the correct list should be [qwen3.5-bf16-mi355x-sglang, qwen3.5-fp8-mi355x-sglang]; (2) pr-link points to PR #980 instead of the current PR #1036; (3) the description says lmsysorg/sglang:v0.5.10-rocm720-mi35x but the actual image in amd-master.yaml is lmsysorg/sglang-rocm:v0.5.10rc0-rocm720-mi35x-20260415 (different repo, RC version, date suffix). Please fix all three before merging.

  Extended reasoning...

  Bug 1 — Duplicate config-key

  The config-keys list in the new perf-changelog entry reads:

  - qwen3.5-fp8-mi355x-sglang   # index 0
  - qwen3.5-bf16-mi355x-sglang  # index 1
  - qwen3.5-fp8-mi355x-sglang   # index 2  ← duplicate

  The FP8 key is present at both index 0 and index 2; the BF16 key is present once. This is a copy-paste artifact — the entry was likely templated from the analogous MI300X/MI325X entry (PR #986) and the first or last line was not changed from FP8 to BF16. The correct list should be [qwen3.5-bf16-mi355x-sglang, qwen3.5-fp8-mi355x-sglang]. If any tooling processes config-keys to build a changelog database, verify coverage, or detect regressions, the duplicate will cause the FP8 config to be processed twice while the BF16 config is silently missing from the record.

  Bug 2 — Wrong PR link

  The entry sets pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/980. This PR is #1036; PR #980 is a completely different, unrelated PR that predates this one by over 50 numbers. Looking at every other entry in perf-changelog.yaml, each pr-link references the specific PR that introduced those changes — the link is the primary audit trail connecting a config change back to a code review. The analogous MI300X/MI325X entry (PR #986) was clearly used as a template and the link was not updated. The correct value is https://github.com/SemiAnalysisAI/InferenceX/pull/1036.

  Bug 3 — Inaccurate image tag in description

  The description bullet reads "Use lmsysorg/sglang:v0.5.10-rocm720-mi35x" but the image actually configured in amd-master.yaml for both qwen3.5-bf16-mi355x-sglang and qwen3.5-fp8-mi355x-sglang is lmsysorg/sglang-rocm:v0.5.10rc0-rocm720-mi35x-20260415. There are three discrepancies: (a) the Docker Hub repository is sglang-rocm not sglang; (b) the version is v0.5.10rc0 (a release candidate), not the stable v0.5.10; (c) the date suffix -20260415 is omitted. For comparison, the analogous MI300X/MI325X entry (PR #986) correctly documents lmsysorg/sglang:v0.5.10-rocm720-mi30x, which matches the actual image used there — the MI355X description appears to have been copied from that entry without updating for the different RC image. The PR description itself also contains this inaccuracy.

  Step-by-step proof

  1. Open perf-changelog.yaml, lines 1362–1369 (new entry added by this PR).
  2. config-keys list: fp8, bf16, fp8 — FP8 duplicated, BF16 appears once. ✗
  3. pr-link: .../pull/980 — current PR is #1036. ✗
  4. Description image: lmsysorg/sglang:v0.5.10-rocm720-mi35x.
  5. Open amd-master.yaml, qwen3.5-bf16-mi355x-sglang entry: image: lmsysorg/sglang-rocm:v0.5.10rc0-rocm720-mi35x-20260415. Mismatch on repo name, RC suffix, and date suffix. ✗
  6. Same image is set for qwen3.5-fp8-mi355x-sglang in the same file. ✗

  All three errors are documentation-only and have no impact on runtime behavior, but they create a misleading changelog entry that should be corrected before merge.

[seungrokj]

lgtm

[chunfangamd]

lgtm