[AMD] Tune Kimi-K2.5 MXFP4 MI355X vLLM toward ATOM parity / 调优 Kimi-K2.5 MXFP4 MI355X vLLM，缩小与 ATOM 的性能差距

[Fangzhou-Ai]

Summary

• Apply the accuracy-gated candidate settings for amd/Kimi-K2.5-MXFP4 on MI355X with vLLM TP4.
• Keep image: TBD intentionally. This draft must not receive a sweep label until an official upstream vLLM ROCm image includes the required AITER MXFP4 backend.
• Preserve the existing TP4 and TP8 search spaces and extend only conc-end from 64 to 128 for both fixed-sequence workloads. Local evidence covers TP4 1k/1k at concurrency 4, 16, 32, and 128; TP8, the remaining points, and 8k/1k remain pending the official-image sweep.
• This changes benchmark configuration only; it does not patch vLLM or AITER source code.
• This draft replaces closed [Superseded] Kimi-K2.5 MI355X tuning draft / [已替代] Kimi-K2.5 MI355X 调优草稿 #2212 after cleaning internal experiment labels from the branch and PR metadata.

Effective settings

VLLM_ROCM_USE_AITER=1
VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1
VLLM_ROCM_USE_SKINNY_GEMM=0
VLLM_ROCM_USE_AITER_RMSNORM=0  # TP4
AITER_MXFP4_INTERMEDIATE=1
AITER_BYPASS_TUNE_CONFIG=0
AITER_MOE_SORT_BACKEND=auto
OMP_NUM_THREADS=1

--kv-cache-dtype fp8
--block-size 16
--max-num-batched-tokens 16384
--max-num-seqs 512
--async-scheduling
--no-enable-prefix-caching

The recipe also preserves the existing 0.90 GPU-memory utilization, MEC firmware workaround, trusted remote code, and multimodal encoder TP mode. It enforces a minimum server context of 9472 tokens so the 1k/1k configuration matches the locally validated server.

Local validation

These results validate local vLLM 94c0ef300 with AITER b2dd7703d, not the pending image:

1k/1k TP4 concurrency	Candidate total tok/s/GPU	ATOM total tok/s/GPU	Candidate / ATOM
4	227.71	246.47	92.39%
16	591.09	640.77	92.25%
32	843.05	918.26	91.81%
128	1736.45	1833.73	94.69%

• GSM8K: 1319/1319 requests, 97.35% strict and flexible exact match.
• bash -n benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh
• Targeted matrix generation: 24 configurations, TP {4,8} × concurrency {4,8,16,32,64,128} × {1k/1k,8k/1k}.

AITER requirement

The tested kernel path requires:

1. ROCm/aiter e7731930 / #3753: transitional FlyDSL 0.2.1 bump.
2. ROCm/aiter 26d40611 / #3764: FlyDSL 0.2.2 compatibility and pin.
3. ROCm/aiter b2dd7703d / #3832: gfx950 RadeonFlow MXFP4 MoE backend and Kimi tuning.

Pinning b2dd7703d or a released descendant is sufficient because it already contains the first two commits. AITER v0.1.17-rc0 contains FlyDSL 0.2.2 but predates #3832, so it cannot reproduce the tested GEMM1/GEMM2 path.

Relationship to #2203

This is a materially different, measured alternative to #2203; only one should merge. #2203 copies K2.6-derived settings such as block size 1, 65,536 batched tokens, AITER RMSNorm, and custom MLA/compilation choices. This draft carries the locally validated Kimi-K2.5 settings that passed throughput scaling and full GSM8K evaluation.

Draft blockers

• Publish and insert an official upstream vLLM ROCm image containing AITER #3832 or a validated successor.
• Finish/update the matching upstream vllm-project/recipes documentation required by InferenceX policy.
• Run full-sweep-fail-fast, including evals, on the final image.
• Coordinate final disposition with Update kimik2.5 fp4 mi355x vLLM config #2203.

AI assistance was used to prepare the benchmark configuration and PR text. The submitter must review and defend every changed line.

中文说明

• 为 MI355X 上的 amd/Kimi-K2.5-MXFP4 vLLM TP4 基准测试应用已通过精度验证的候选配置。
• 有意保留 image: TBD。在上游正式发布包含所需 AITER MXFP4 后端的 vLLM ROCm 镜像之前，本草稿 PR 不应添加扫描标签。
• 保留现有 TP4 和 TP8 搜索空间，仅将两个固定序列长度负载的 conc-end 从 64 扩展到 128。本地数据覆盖 TP4 1k/1k 的并发点 4、16、32 和 128；TP8、其余并发点与 8k/1k 需要在正式镜像发布后通过 CI 扫描验证。
• 本 PR 只修改基准测试配置，不修改 vLLM 或 AITER 源码。
• 本草稿用于替代已关闭的 [Superseded] Kimi-K2.5 MI355X tuning draft / [已替代] Kimi-K2.5 MI355X 调优草稿 #2212，并清理了分支及 PR 元数据中的内部实验标签。

“Effective settings”代码块列出了完整启动配置。该配置还保留现有的 0.90 GPU 显存利用率、MEC 固件规避措施、trust-remote-code 和多模态编码器 TP 模式，并将服务端最小上下文长度设为 9472，以对齐本地验证的 1k/1k 配置。

本地验证

上表数据来自本地 vLLM 94c0ef300 与 AITER b2dd7703d，不代表待发布镜像的结果：候选配置在并发 4、16、32、128 下达到 ATOM 吞吐量的 91.81%–94.69%。GSM8K 共完成 1319/1319 个请求，严格和宽松 exact match 均为 97.35%。Shell 语法检查通过；定向矩阵生成了 24 个配置，即 TP {4,8} × 并发 {4,8,16,32,64,128} × {1k/1k,8k/1k}。

AITER 依赖

所测内核路径需要 ROCm/aiter e7731930（#3753，过渡到 FlyDSL 0.2.1）、26d40611（#3764，兼容并固定 FlyDSL 0.2.2）以及 b2dd7703d（#3832，为 gfx950 增加 RadeonFlow MXFP4 MoE 后端和 Kimi 调优）。直接使用 b2dd7703d 或包含它的后续正式版本即可，因为该提交已经包含前两项。

AITER v0.1.17-rc0 虽然包含 FlyDSL 0.2.2，但发布时间早于 #3832，因此无法复现本次验证所用的 GEMM1/GEMM2 路径。

与 #2203 的关系

本 PR 是与 #2203 实质不同、具有实测依据的替代方案，两者最终只能合并一个。#2203 复用了 K2.6 配置，包括 block size 1、65,536 batched tokens、AITER RMSNorm 以及自定义 MLA/编译选项；本草稿采用已完成吞吐量扩展测试和完整 GSM8K 评估的 Kimi-K2.5 候选配置。

草稿阻塞项

• 发布并填入包含 AITER #3832 或已验证后续版本的上游正式 vLLM ROCm 镜像。
• 按 InferenceX 要求完成或更新对应的上游 vllm-project/recipes 文档。
• 使用最终镜像运行包含精度评估的 full-sweep-fail-fast。
• 与 Update kimik2.5 fp4 mi355x vLLM config #2203 协调最终处理方式。

本 PR 的基准配置和说明文本使用了 AI 辅助。提交者必须逐行审核，并能够完整解释和维护这些改动。

[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=29854639662
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29854639662

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29855213440
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29855213440

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29855213440
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29855213440

[chunfangamd]

vllm-project/recipes#660

[chunfangamd]

@claude check the recipe RP: vllm-project/recipes#673

[Klaud-Cold]

Claude finished @chunfangamd's task in 1m 19s —— View job

---

Recipe PR check — vllm-project/recipes#673

• Read vllm-project/recipes#673
• Compare against this PR's amd/Kimi-K2.5-MXFP4 requirement (Check 3)
• Report verdict

✅ This resolves the Check 3 blocker

Recipe #673 — "[ROCm] Document Kimi-K2.5 MXFP4 MI355X benchmark reproduction" — is MERGED (merged 2026-07-23T08:19:11Z by @xianbaoqian, merge commit 0885682), and unlike the previously-linked #660 it documents the correct model.

Item	Status
Merged into main	✅ MERGED
File touched	models/moonshotai/Kimi-K2.5.yaml (+56/−4)
Model / model-path	✅ amd/Kimi-K2.5-MXFP4 — matches this PR
Hardware / lane	✅ MI355X, TP4 high-concurrency benchmark lane

This is the fix for the earlier Klaud-Cold rejection, which failed Check 3 solely because #660 documented amd/MiniMax-M3-MXFP4. #673 adds a dedicated "MXFP4 benchmark reproduction (InferenceX MI355X sweep)" section to the Kimi-K2.5 recipe.

Settings cross-check (recipe vs. this PR's config)

The recipe's reproduction section matches this PR's effective settings:

• Env vars all present and equal: VLLM_ROCM_USE_AITER=1, VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4, VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1, VLLM_ROCM_USE_SKINNY_GEMM=0, VLLM_ROCM_USE_AITER_RMSNORM=0, AITER_MXFP4_INTERMEDIATE=1, AITER_BYPASS_TUNE_CONFIG=0, AITER_MOE_SORT_BACKEND=auto, OMP_NUM_THREADS=1. ✅
• Serve args match: --tensor-parallel-size 4, --kv-cache-dtype fp8, --block-size 16, --max-num-batched-tokens 16384, --max-num-seqs 512, --async-scheduling, --no-enable-prefix-caching, --mm-encoder-tp-mode data. ✅
• MAX_MODEL_LEN >= 9472 guard documented, matching the 1k/1k validated server. ✅
• Uses upstream vllm/vllm-openai-rocm:nightly. ✅

One minor discrepancy (non-blocking)

The merged recipe reproduction command shows --gpu-memory-utilization 0.85, whereas this PR's body and config state 0.90. This is not a Check-3 concern (model/model-path is the MAJOR arg and matches), but for exact reproduction fidelity you may want to align the recipe's gpu-memory-utilization with the benchmarked 0.90, or note why the sweep uses 0.90 while the recipe documents 0.85.

Bottom line: #673 satisfies the merged-upstream-recipe requirement for amd/Kimi-K2.5-MXFP4 that #660 did not. The draft blocker "Finish/update the matching upstream recipes documentation" can now be checked, and a re-run of the Klaud-Cold gate should clear Check 3. Worth updating the linked recipe in the PR checklist from #660 → #673.
· branch perf/kimi-k2.5-mi355x-tp4

[chunfangamd]

@claude I approve this PR because the reported issue has been resolved.

Signed: @chunfangamd

[Klaud-Cold]

Claude finished @chunfangamd's task in 1m 42s —— View job

---

Approval acknowledged — one step still needed to flip the red check

Thanks @chunfangamd. I've verified the resolution, and the underlying blocker is resolved — but the codeowner-signoff-verify status check is still red, and a comment approval alone won't turn it green. Here's the exact state.

• Fetch base branch and inspect current PR diff
• Verify recipe [AMD] Add ATOM+MTP, dsr1 fp8/fp4 #673 resolves the Check 3 blocker (correct model)
• Check whether the PR checklist still references the stale recipe Enable fast mode for Claude Code workflows #660
• Report status back to @chunfangamd

What's resolved ✅

• Recipe (Check 3): vllm-project/recipes#673 is MERGED (merge commit 0885682, by @xianbaoqian) and documents the correct model — models/moonshotai/Kimi-K2.5.yaml for amd/Kimi-K2.5-MXFP4. This is the fix for the earlier Klaud-Cold rejection, which failed only because Enable fast mode for Claude Code workflows #660 documented amd/MiniMax-M3-MXFP4.
• Official image: configs/amd-master.yaml now pins the upstream vllm/vllm-openai-rocm:nightly-387189c… build (no more image: TBD), satisfying the engine-first / upstream-image requirement.
• Settings ↔ recipe match: the script's effective settings (--kv-cache-dtype fp8, --block-size 16, --max-num-batched-tokens 16384, --max-num-seqs 512, --async-scheduling, --gpu-memory-utilization 0.85, the AITER env vars, and the MAX_MODEL_LEN >= 9472 guard) line up with the merged recipe. Note the earlier "0.90 vs 0.85 discrepancy" is now gone — commit 19100b1 lowered util to 0.85, matching [AMD] Add ATOM+MTP, dsr1 fp8/fp4 #673.

What still blocks the green check ⚠️

The commit status codeowner-signoff-verify on head 45c6584 is failure, pointing at the Klaud-Cold rejection from 2026-07-23 05:40 — that verdict was computed before #673 merged (08:19), and against your APPROVED checklist whose Additional detail section still links the wrong recipe:

### Additional detail section:
https://github.com/vllm-project/recipes/pull/660   ← MiniMax-M3, stale

To clear the gate:

1. Update the sign-off checklist so the Additional detail section links [AMD] Add ATOM+MTP, dsr1 fp8/fp4 #673 instead of Enable fast mode for Claude Code workflows #660, and check the draft blocker "Finish/update the matching upstream recipes documentation." Fix this →
2. Re-run the sign-off verification (e.g. re-post /reuse-sweep-run or re-trigger the Klaud-Cold gate) so it re-evaluates against the now-merged [AMD] Add ATOM+MTP, dsr1 fp8/fp4 #673. Check 3 should then pass, and with Checks 0–2 and 4–9 already green the status should flip to success.

One note on the PR body

The PR description still reads image: TBD and cites --block-size 16 … 0.90 / 1k/1k local validation, but the branch has since pinned the official nightly, lowered util to 0.85, and (per perf-changelog.yaml) deprecated 1k1k in favor of the 8k1k sweep (#2263). Worth refreshing the body so it matches the merged config — cosmetic, not a merge blocker.

I can't approve the PR myself (that's disabled for the bot for security reasons), but once the checklist link is corrected and the sign-off gate re-runs, this should be green and mergeable.
· branch perf/kimi-k2.5-mi355x-tp4

[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/29953041481
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. Range: 96.89%-97.42%, link: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29953041481
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
• 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 every single-node vLLM/SGLang recipe in this PR is documented in the official vLLM recipes and/or the SGLang cookbook:
  • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
• Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked 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
  [ROCm] Document Kimi-K2.5 MXFP4 MI355X benchmark reproduction vllm-project/recipes#673
  Signed: @chunfangamd

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

Supersedes the earlier REJECTED verdict on this SHA: that gate ran before vllm-project/recipes#673 merged and against the stale #660 link. The current sign-off links #673 (now MERGED), which documents the correct model, so Check 3 now passes. Green run: 29953041481.

✅ Check 0 (CODEOWNER): PASS — @chunfangamd owns configs/amd-master.yaml; the script + perf-changelog.yaml are catch-all (* @InferenceX/core), covered by any CODEOWNER.
✅ Check 1 (green in-PR sweep+eval): PASS — head 45c6584 has executed, green single-node 8k1k / (TP4/TP8) and per-config eval / jobs in run 29953041481 (1k1k skip is expected — config sweeps 8k1k only).
✅ Check 2 (eval accuracy): PASS — GSM8K em_strict 96.89%–97.42% (n=1319) on amd/Kimi-K2.5-MXFP4, same image vllm/vllm-openai-rocm:nightly-387189c… as this PR's config.
✅ Check 3 (recipe merged+complete): PASS — recipes#673 is MERGED (2026-07-23) and documents amd/Kimi-K2.5-MXFP4 (MI355X TP4 sweep). Major args match: model, TP4, kv-cache-dtype fp8, block-size 16, batched-tokens 16384, max-num-seqs 512, async-scheduling, gpu-mem-util 0.85, and the AITER MXFP4 env vars.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by @chunfangamd (COLLABORATOR).
✅ Check 5 (latest template): PASS — all current-template items present; the two unchecked items (agentic spec-decode, catch-all reasoning) are non-applicable conditionals.
✅ Check 6 (upstream image / engine-first): PASS — vllm/vllm-openai-rocm:nightly-387189c… is the upstream vLLM org; framework vllm on established SKU MI355X. Engine-first N/A (vLLM entry).
✅ Check 7 (no arch hacks): PASS — no --hf-overrides; fp8 KV-cache, block-size, batched-tokens are serving config, not model-FLOPs reduction.
➖ Check 8 (spec-decode chat template): N/A — no speculative-decoding changes.
✅ Check 9 (no engine patch): PASS — env-var exports and serve args only; pip install amd-quark is a stock upstream quantization dependency, not an engine fork/patch.
➖ Check 10 (agentic golden AL): N/A — no agentic speculative-decoding changes.

[Klaud-Cold]

❌❌❌ REJECTED ❌❌❌

@chunfangamd — blocking: this sign-off's Additional detail section links vllm-project/recipes#660, which documents the wrong model (amd/MiniMax-M3-MXFP4, models/MiniMaxAI/MiniMax-M3.yaml), not this PR's amd/Kimi-K2.5-MXFP4. The correct merged recipe #673 (models/moonshotai/Kimi-K2.5.yaml) is only linked in a later comment. Update this sign-off's Additional detail section to link #673, then re-run the gate.

✅ Check 0 (CODEOWNER): PASS — @chunfangamd owns configs/amd-master.yaml; the script + perf-changelog.yaml are catch-all (* @InferenceX/core), covered by any CODEOWNER.
✅ Check 1 (green in-PR sweep+eval): PASS — in-PR commit 45c6584 has executed, green single-node 8k1k / (TP4+TP8) and per-config eval / jobs in run 29953041481. (Head d1fd00e is a main-merge commit with all sweep jobs skipped; validation keys off any in-PR commit.)
✅ Check 2 (eval accuracy): PASS — GSM8K em_strict 96.89%–97.42% (n=1319) on amd/Kimi-K2.5-MXFP4, same image vllm/vllm-openai-rocm:nightly-387189c… as this PR's config.
❌ Check 3 (recipe merged+complete): FAIL — sign-off links #660 (MERGED, but documents amd/MiniMax-M3-MXFP4 with --tool-call-parser minimax_m3); model/model-path is a MAJOR arg and mismatches this PR's amd/Kimi-K2.5-MXFP4. The correct merged recipe #673 (Kimi-K2.5, merged 2026-07-23) is not linked in this sign-off — only in a separate comment. Link #673 in the Additional detail section.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by @chunfangamd (COLLABORATOR).
✅ Check 5 (latest template): PASS — all current-template items present; the two unchecked items (agentic spec-decode, catch-all reasoning) are non-applicable conditionals.
✅ Check 6 (upstream image / engine-first): PASS — vllm/vllm-openai-rocm:nightly-387189c… is the upstream vLLM org; framework: vllm, single-node established SKU MI355X. Engine-first N/A (vLLM entry).
✅ Check 7 (no arch hacks): PASS — no --hf-overrides; fp8 KV-cache / block-size / batched-tokens are serving config, not model-FLOPs reduction.
➖ Check 8 (spec-decode chat template): N/A — no speculative-decoding changes.
✅ Check 9 (no engine patch): PASS — env-var exports and serve args only; pip install amd-quark is a stock upstream quantization dependency, not an engine fork/patch.
➖ Check 10 (agentic golden AL): N/A — no agentic speculative-decoding changes.

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

Supersedes the earlier REJECTED verdict on this SHA: that gate ran against a sign-off that linked the wrong recipe (#660, MiniMax-M3). The current sign-off links vllm-project/recipes#673 (MERGED, correct model), so Check 3 now passes. Green run: 29953041481.

✅ Check 0 (CODEOWNER): PASS — @chunfangamd owns configs/amd-master.yaml; the script + perf-changelog.yaml are catch-all (* @InferenceX/core), covered by any CODEOWNER.
✅ Check 1 (green in-PR sweep+eval): PASS — in-PR commit 45c6584 has executed, green single-node 8k1k / (TP4+TP8) and per-config eval / jobs in run 29953041481. (1k1k skip is expected — the config sweeps 8k1k only; head d1fd00e is a main-merge commit whose jobs are skipped, but validation keys off any in-PR commit.)
✅ Check 2 (eval accuracy): PASS — GSM8K em_strict 96.89% (TP8 c64) – 97.12% (TP4 c128), n=1319, on amd/Kimi-K2.5-MXFP4, same image vllm/vllm-openai-rocm:nightly-387189c… as this PR's config.
✅ Check 3 (recipe merged+complete): PASS — recipes#673 MERGED 2026-07-23, documents amd/Kimi-K2.5-MXFP4 (models/moonshotai/Kimi-K2.5.yaml) with a dedicated MI355X sweep-reproduction section. Major args match: model, TP4, --kv-cache-dtype fp8, --block-size 16, --max-num-batched-tokens 16384, --max-num-seqs 512, --async-scheduling, --gpu-memory-utilization 0.85, and the AITER MXFP4 env vars.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by @chunfangamd (COLLABORATOR).
✅ Check 5 (latest template): PASS — all current-template items present; the two unchecked items (agentic spec-decode, catch-all reasoning) are non-applicable conditionals.
✅ Check 6 (upstream image / engine-first): PASS — vllm/vllm-openai-rocm:nightly-387189c… is the upstream vLLM org; framework: vllm, single-node established SKU MI355X. Engine-first N/A (vLLM entry).
✅ Check 7 (no arch hacks): PASS — no --hf-overrides; fp8 KV-cache / block-size / batched-tokens are serving config, not model-FLOPs reduction.
➖ Check 8 (spec-decode chat template): N/A — no speculative-decoding changes.
✅ Check 9 (no engine patch): PASS — env-var exports and serve args only; pip install amd-quark is a stock upstream MXFP4 quantization dependency, not an engine fork/patch.
➖ Check 10 (agentic golden AL): N/A — no agentic speculative-decoding changes (this PR touches the non-agentic kimik2.5-fp4-mi355x-vllm entry only).

[functionstackx]

lgtm besides this error here in recipes, can u please fix vllm-project/recipes#673 (comment)

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

Green run: 29953041481. Head 1f924f9 matches the signed-off SHA (a main-merge commit); the swept commit 45c6584 is still in the PR and its kimik2.5_fp4_mi355x.sh and kimik2.5-fp4-mi355x-vllm config entry are byte-identical to the head, so the validated recipe is the one being merged.

✅ Check 0 (CODEOWNER): PASS — @chunfangamd is listed on configs/amd-master.yaml; the script and perf-changelog.yaml fall to the catch-all * @InferenceX/core, covered by any CODEOWNER.
✅ Check 1 (green in-PR sweep+eval): PASS — in-PR commit 45c6584 has executed, success single-node 8k1k / (TP4 c4–c128, TP8 c8–c128) and per-config eval / check-runs in run 29953041481. The single-node 1k1k / skip is expected — the config declares only the 8k1k scenario.
✅ Check 2 (eval accuracy): PASS — GSM8K em_strict 96.89%–97.42% (n=1319) across all four eval arms (TP4 c64/c128, TP8 c64/c128) on amd/Kimi-K2.5-MXFP4, run on vllm/vllm-openai-rocm:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 — the same image this PR pins. In line with the ~0.9697 upstream recipe reference.
✅ Check 3 (recipe merged+complete): PASS — vllm-project/recipes#673 is MERGED (2026-07-23) and its "MXFP4 benchmark reproduction (InferenceX MI355X sweep)" section in models/moonshotai/Kimi-K2.5.yaml documents this exact lane. Major args all match: model amd/Kimi-K2.5-MXFP4, MI355X, TP4 (recipe default covers the TP8 arm), --kv-cache-dtype fp8, --block-size 16, --max-num-batched-tokens 16384, --max-num-seqs 512, --async-scheduling, --no-enable-prefix-caching, --mm-encoder-tp-mode data, --gpu-memory-utilization 0.85, upstream vllm/vllm-openai-rocm:nightly, and every AITER env var including the TP<8 VLLM_ROCM_USE_AITER_RMSNORM=0 guard. Informational only: the commit-pinned nightly tag and sweep concurrency ranges are InferenceX harness specifics.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by @chunfangamd (COLLABORATOR).
✅ Check 5 (latest checklist): PASS — every item in the current template is present and checked; the two unchecked items are the non-applicable conditionals (agentic spec-decode, and the catch-all "additional reasoning" clause).
✅ Check 6 (upstream image / engine-first): PASS — vllm/vllm-openai-rocm:nightly-387189c… is the upstream vLLM Docker Hub org; framework: vllm on established SKU MI355X. Engine-first ordering N/A — this modifies an existing vLLM entry and adds no other framework.
✅ Check 7 (no arch hacks): PASS — no --hf-overrides or config rewriting; fp8 KV cache, block size, batched-token and sequence caps are serving/scheduler config, and the fp8 KV precision is evals-gated at ~97% GSM8K.
➖ Check 8 (spec-decode chat template): N/A — this PR adds no speculative-decoding config.
✅ Check 9 (no engine patch): PASS — the diff only adds env-var exports, a MAX_MODEL_LEN >= 9472 floor, and serve flags. No .patch/git apply, heredoc rewrite of engine sources, site-packages edit, or forked engine wheel. The pre-existing pip install amd-quark is a stock upstream PyPI quantization dependency (vllm#35633), not an engine fork.
➖ Check 10 (agentic golden AL): N/A — no agentic speculative-decoding changes; this PR touches only the non-agentic kimik2.5-fp4-mi355x-vllm entry.

[chunfangamd]

lgtm besides this error here in recipes, can u please fix vllm-project/recipes#673 (comment)

vllm-project/recipes#680

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

Green run: 29953041481. Head 4c45983 is the signed-off SHA (a main-merge commit); the swept commit 45c6584 is still in the PR and its kimik2.5_fp4_mi355x.sh and kimik2.5-fp4-mi355x-vllm entry are identical to the head, so the validated recipe is the one being merged.

✅ Check 0 (CODEOWNER): PASS — @chunfangamd is listed on configs/amd-master.yaml; the script and perf-changelog.yaml fall to the catch-all * @InferenceX/core, covered by any CODEOWNER.
✅ Check 1 (green in-PR sweep+eval): PASS — in-PR commit 45c6584 has executed, success single-node 8k1k / (TP4 c4–c128, TP8 c8–c128) and per-config eval / check-runs in run 29953041481. The single-node 1k1k / skip is expected — the config declares only the 8k1k scenario.
✅ Check 2 (eval accuracy): PASS — GSM8K em_strict 96.89%–97.42% (n=1319) across all four eval arms (TP4 c64/c128, TP8 c64/c128) on amd/Kimi-K2.5-MXFP4, run on vllm/vllm-openai-rocm:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 — the same image this PR pins.
✅ Check 3 (recipe merged+complete): PASS — vllm-project/recipes#673 is MERGED (2026-07-23) and documents amd/Kimi-K2.5-MXFP4 on MI355X. Major args match: model, TP4, --kv-cache-dtype fp8, --block-size 16, --max-num-batched-tokens 16384, --max-num-seqs 512, --async-scheduling, --no-enable-prefix-caching, --mm-encoder-tp-mode data, gpu-mem-util 0.85, and the AITER/INT4 env vars. Informational only: the script gates VLLM_ROCM_USE_AITER_RMSNORM=0 on TP<8 where the recipe exports it unconditionally, and the image tag differs from the recipe's generic nightly.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by @chunfangamd (COLLABORATOR).
✅ Check 5 (latest checklist): PASS — every item of the current PR_REVIEW_CHECKLIST.md template is present and checked, except the agentic spec-decode item, which is inapplicable here (see Check 10).
✅ Check 6 (upstream image / engine-first): PASS — kimik2.5-fp4-mi355x-vllm runs vllm/vllm-openai-rocm:nightly-387189c… from the upstream vLLM org; framework: vllm, so engine-first ordering does not apply.
✅ Check 7 (no architecture hacks): PASS — no --hf-overrides / override args; the changes are scheduler, block-size, memory and KV-cache-precision knobs, and FP8 KV cache is covered by the passing evals.
➖ Check 8 (spec-decode chat template): N/A — no speculative-decoding changes in this PR.
✅ Check 9 (no engine patches): PASS — no .patch, git apply, heredoc/sed rewrites of engine sources, site-packages edits, or forked engine wheels. Informational only: the pre-existing pip install amd-quark (already on main, not added here) pulls the official AMD PyPI quantization package to work around vllm-project/vllm#35633 — not a fork or rebuild of the engine.
➖ Check 10 (agentic golden AL): N/A — this PR touches only the non-agentic fixed_seq_len lane and adds no speculative-decoding config.

[chunfangamd]

https://github.com/vllm-project/recipes/pull/703/changes

[Fangzhou-Ai]

Hi @functionstackx the recipe has been updated to fix the missing envs (vllm-project/recipes#703), can we merge the PR now?

[functionstackx]

/reuse-sweep-run