config(dsv41flash): re-sweep MI355X on the first ROCm 10.0 nightly carrying vllm#58510 / 在首个包含 vllm#58510 的 ROCm 10.0 nightly 上重新扫描 MI355X - #3420
Conversation
Repin dsv41flash-fp4-mi355x-vllm-agentic-dspark off nightly-rocm100-7f1a5398, which PR #3404 qualified yesterday, and re-measure all 16 points on the next nightly. The image tag is TBD until that build publishes; this PR is not dispatchable until it is replaced. Nothing else changes: the TP=4 and TP=2 rows keep concurrency 1-128 and the recipe script is untouched. 将 dsv41flash-fp4-mi355x-vllm-agentic-dspark 从 PR #3404 昨日验证的 nightly-rocm100-7f1a5398 迁出，在下一个 nightly 上重新测量全部 16 个数据点。 镜像标签暂为 TBD，待该构建发布后填入；在此之前本 PR 不可派发。其余内容不变： TP=4 与 TP=2 两行保持并发 1-128，recipe 脚本未作改动。 Co-authored-by: Cursor <cursoragent@cursor.com>
|
Thanks for the contribution!
中文感谢你的贡献！
|
Replace the XXX placeholder now that the draft PR exists. 在草稿 PR 建立后，将 changelog 条目中的 XXX 占位符替换为实际链接。 Co-authored-by: Cursor <cursoragent@cursor.com>
The target image is the first nightly whose commit contains that merge, not simply the next nightly published. 将本次重新扫描所用镜像的选取条件限定为包含 vllm-project/vllm#58510 的第一个 nightly，而非简单地取下一个发布的 nightly。 Co-authored-by: Cursor <cursoragent@cursor.com>
Fangzhou-Ai
changed the title
First ROCm 10.0 nightly containing vllm-project/vllm#58510. 首个包含 vllm-project/vllm#58510 的 ROCm 10.0 nightly。 Co-authored-by: Cursor <cursoragent@cursor.com>
|
Claude encountered an error after 0s —— View job I'll analyze this and get back to you. |
1 similar comment
|
Claude encountered an error after 0s —— View job I'll analyze this and get back to you. |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.
Beyond the inline finding, I checked a few other conventions this recipe could implicate: the "model.container must equal image" rule doesn't apply since this recipe is single-node (multinode: false, no launcher/container field); the changelog's cited "golden acceptance 3.51" matches the real thinking_on/draft-length-5 entry in golden_al_distribution/dsv41flash_dspark.yaml; and the draft-precision env-var checklist item is vLLM/SGLang-specific tooling that doesn't apply here since this recipe's framework is vllm.
Extended reasoning...
Config-only change: an image tag bump in configs/amd-master.yaml plus a matching append-only perf-changelog.yaml entry, no code or script changes. No security-sensitive surface (auth, crypto, permissions) is touched. One confirmed inline finding (stale "tag TBD" prose in the changelog contradicting the already-pinned concrete tag) is being posted separately, so a human look is warranted; I additionally checked and ruled out the multi-node container-equals-image rule, the golden acceptance-length provenance, and the draft-precision env-var concern as inapplicable to this diff.
| - "The pin is mainly for vllm-project/vllm#58510, which computes the MXFP8 GEMM on native 32x32 block scales for gfx950. This arm serves a checkpoint whose non-routed weights are MXFP8 on exactly that architecture, so the target image is the first nightly whose commit contains that merge rather than simply the next one published. The exact tag, its publication time and digest are filled in once that build exists." | ||
| - "在下一个 ROCm 10.0 nightly 上重新扫描 MI355X DeepSeek-V4.1-Flash vLLM AgentX 臂，替换 nightly-rocm100-7f1a5398。本次仅更换镜像：TP=4 与 TP=2 两行保持并发 1、2、4、8、16、32、64 与 128，其余 recipe 设置全部不变，包括按 TP 与并发决定的 Engram 放置、prefill 分块与序列上限阶梯、图捕获，以及吞吐用黄金接受长度 3.51、评测用真实分块验证的五 token DSpark。" |
There was a problem hiding this comment.
🟡 (optional) Maintainers reading this changelog entry are told the pin is still unresolved, even though configs/amd-master.yaml:1330 already ships a concrete tag. The entry (perf-changelog.yaml:8897-8898) says "The exact tag, its publication time and digest are filled in once that build exists," but commit fe94952 already replaced the image: TBD placeholder with nightly-rocm100-29468dde8b515031dc6d4d9d06bf0a2fa0442098 without updating this text. Since perf-changelog.yaml is append-only and historical entries must never be edited, this stale placeholder becomes permanent, so anyone auditing this PR later cannot tell from the changelog which tag was actually swept or verify its ancestry/digest claim. …
Why this was flagged
…Fix: when the tag is filled in, append a follow-up changelog entry (not edit the existing one) recording the actual tag, its publication time, and digest so the audit trail matches configs/amd-master.yaml.
The trigger is commit fe94952 filling in configs/amd-master.yaml:1330 with a concrete image tag after perf-changelog.yaml's entry (added earlier in the same PR, lines 8889-8898) was written describing the tag as still-to-be-filled. The changelog text at perf-changelog.yaml:8898 was never updated or supplemented to reflect the resolved tag. On the base branch there was no such entry at all; after merge, the permanent append-only record for PR #3420 misdescribes the current state of the pin it documents. No safeguard catches this because perf-changelog.yaml correctness isn't validated against configs/amd-master.yaml's actual image value.
Verification: nit. The discrepancy is real. On HEAD, configs/amd-master.yaml:1330 ships a concrete pin image: vllm/vllm-openai-rocm:nightly-rocm100-29468dde8b515031dc6d4d9d06bf0a2fa0442098, while the PR's new changelog entry still states the tag is unresolved: perf-changelog.yaml:8897 (English) "The exact tag, its publication time and digest are filled in once that build exists" and the Chinese…
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=36101751233 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=36101751233 |
|
/reuse-sweep-run 36101751233 |
…ng vllm#58510 Move dsv41flash-fp4-mi355x-vllm-agentic-dspark off nightly-rocm100-7f1a5398 to the first nightly-rocm100 build whose commit contains vllm-project/vllm#58510 (MXFP8 GEMM on native 32x32 block scales for gfx950). Only the image tag changes; TP=4/TP=2 rows keep concurrency 1,2,4,8,16,32,64,128 and the benchmark script is untouched. Tag stays TBD until #58510 merges and a carrying nightly publishes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
seungrokj
left a comment
seungrokj
left a comment
There was a problem hiding this comment.
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.
- Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
- Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
- Verified that every draft model and draft head is served as it ships: the draft that ships with the served checkpoint, at its stored precision, through the pinned upstream image's default handling, with the shipped and effective draft precision recorded in the additional detail section. No submission-side quantization, dtype override, checkpoint substitution, or patch may lower draft precision below that default, regardless of eval results or AL. Explicitly verified that
SGLANG_NVFP4_CKPT_FP8_NEXTN_MOEis not enabled in the effective recipe, including inherited settings; enabling it is prohibited going forward, and historical runs do not grant an exception. See Draft-model precision for what counts as the default and the MLPerf comparison. - 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 against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
- 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; target/verifier FLOPs at lower precisions is fine, given that the config passes private evals, but this does not permit lowering draft-model or draft-head precision below what ships. 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 this PR uses
append-only: true, verified that it only adds generated points or recipe variants inside a selected existing config/scenario and existing same-image visual curve: every previously generated point remains present with the same recipe, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it. - If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
- Reported measured throughput/E2EL Pareto counts and evidence per affected curve (≥5 points strongly recommended). Below 5 or unverifiable: tag a core maintainer for review; recorded admin bypass required before merge. N/A if no curves are affected. Details.
Additional detail section:
- insert any additional info here
- recipe at https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash
- vllm does not use any quantization on the draft layers of dspark layers of deepseek-ai/DeepSeek-V4.1-Flash model. relevant codes are https://github.com/SemiAnalysisAI/InferenceX/blob/config/dsv41flash-mi355x-resweep-nightly/benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh#L121-L148
Signed: seungrokj
✅✅✅ Verdict: PASS ✅✅✅Assessed at the signed-off SHA Passed and not applicable checks✅ Check 0 (CODEOWNER): PASS — ✅ Check 1 (Sweep on in-PR commit): PASS — In-PR commit ✅ Check 2 (Evals): PASS — ✅ Check 3 (Recipe): PASS — The linked published recipe (vllm-project/recipes ✅ Check 4 (Reuse command): PASS — ✅ Check 5 (Latest template): PASS — Every item in the current ✅ Check 6 (Upstream image / ordering): PASS — ✅ Check 7 (Deprecated models): PASS — ✅ Check 8 (No arch hacks): PASS — This PR changes only the image. The effective script has no ✅ Check 9 (Chat template): PASS — The agentic replay uses ✅ Check 10 (No engine patches): PASS — Nothing patches the engine. The only pip installs are pre-existing eval/harness dependencies. ✅ Check 11 (Golden AL): PASS — Throughput mode sets ➖ Check 12 (Append-only): N/A — The new changelog entry has no ✅ Check 13 (Draft as shipped): PASS — The embedded DSpark head ( ✅ Check 14 (Pareto coverage): PASS — There is one curve: dsv41flash / agentic-coding / MI355X vLLM FP4 / Assessed commit: |
3 participants
Description
Re-sweeps
dsv41flash-fp4-mi355x-vllm-agentic-dsparkon the next ROCm 10.0 nightly, moving offnightly-rocm100-7f1a5398e9610d96c473931a26c0e12bbe0d0423which #3404 qualified yesterday.The image tag is deliberately
TBDwhile this PR is a draft. The target is not simply the next nightly: it is the firstnightly-rocm100build whose commit contains vllm-project/vllm#58510, MXFP8 GEMM on native 32x32 block scales for gfx950. That PR is still open. Once it merges and a nightly carrying it publishes, I will fill in the tag, record its digest, take the PR out of draft and add the sweep label in one follow-up push. Do not dispatch until then.Why this pin waits on #58510
This arm runs gfx950 and serves a checkpoint whose non-routed weights are MXFP8, so #58510 lands directly on its hot path; it also touches
rocm_aiter_mla_sparse.py, which this recipe's sparse decode uses. Pinning whichever nightly happens to publish first would produce a curve that does not measure the change the re-sweep exists for. Membership is therefore checked by commit ancestry against the PR's merge commit, not by publication date, and a nightly that arrives before the merge is skipped.Only the
image:line changes. The TP=4 and TP=2 rows keep concurrency 1, 2, 4, 8, 16, 32, 64 and 128, andbenchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.shis untouched: Engram placement per TP and concurrency, the prefill-chunk and sequence-cap ladder, graph capture,--no-swa-bounded-replay, and five-token DSpark with golden acceptance 3.51 for throughput and real block verification for evals all stay as they are onmain.The MI300X and MI325X arms are not touched.
Validation so far
configs/amd-master.yaml,configs/runners.yamlandperf-changelog.yamlall parse andgit diff --checkis clean. Targeted generation emits exactly the expected 16 points:They currently carry
"image": "TBD". Re-run this after the tag lands and confirm no placeholder survives. GPU sweep and eval results are pending.AI model disclosure
claude-opus-5-thinking-high), via the Cursor agent.Related Issue
N/A
Type of Change
Checklist
perf-changelog.yamland have not edited historical entriesOWNER/MEMBER/COLLABORATOR) has commented/use <run_id>(or the legacy/reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps. Remove and re-add the label to force one.中文
说明
在下一个 ROCm 10.0 nightly 上重新扫描
dsv41flash-fp4-mi355x-vllm-agentic-dspark，迁出昨日由 #3404 完成资格验证的nightly-rocm100-7f1a5398e9610d96c473931a26c0e12bbe0d0423。在本 PR 处于草稿状态期间，镜像标签特意保留为
TBD。 目标并非简单地取下一个 nightly，而是第一个其提交包含 vllm-project/vllm#58510（在 gfx950 上使用原生 32x32 block scale 计算 MXFP8 GEMM）的nightly-rocm100构建。该 PR 尚未合并。合并且包含它的 nightly 发布后，我会在一次后续推送中填入标签、记录其摘要、取消草稿状态并添加 sweep 标签。在此之前请勿派发。为何本次固定要等 #58510
本臂运行于 gfx950，所服务的 checkpoint 其非 routed 权重为 MXFP8，因此 #58510 直接作用于其热路径；该 PR 同时修改了本 recipe 稀疏解码所用的
rocm_aiter_mla_sparse.py。若随意选取最先发布的 nightly，得到的曲线并未测量本次重扫所针对的改动。因此此处通过与该 PR 合并提交的祖先关系来判定包含性，而非发布时间；在合并之前发布的 nightly 会被跳过。本次仅改动
image:一行。TP=4 与 TP=2 两行保持并发 1、2、4、8、16、32、64 与 128，且benchmarks/single_node/agentic/dsv41flash_fp4_mi355x_vllm_mtp.sh未作改动：按 TP 与并发决定的 Engram 放置、prefill 分块与序列上限阶梯、图捕获、--no-swa-bounded-replay，以及吞吐用黄金接受长度 3.51、评测用真实分块验证的五 token DSpark，均与main上保持一致。MI300X 与 MI325X 两个臂不受影响。
目前的验证
configs/amd-master.yaml、configs/runners.yaml与perf-changelog.yaml均可解析，git diff --check无告警。定向生成恰好产出预期的 16 个数据点，命令与结果见上方英文部分。这些点目前携带
"image": "TBD"。标签落定后需重新运行该命令并确认不再残留占位符。GPU sweep 与评测结果尚待完成。AI 模型使用说明、关联 issue、改动类型与检查清单见上方英文部分。