[AMD][MI35X] 0907 DSV4 sglang mtp 8k1k - #2910
Conversation
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Hi 1am9trash — I reviewed this PR and didn't find any bugs. Because it touches a CODEOWNER-owned config file (configs/amd-master.yaml) and an AMD MI355X production recipe, a human/CODEOWNER look is still worthwhile per the repo's review process.
What was reviewed:
- The image bump in configs/amd-master.yaml against the dated-nightly tag pattern other sglang-rocm entries follow
- The SHARED_EXPERTS_ARGS rework (default enforce-fusion, disable only when EP_SIZE>1) and the two new exported env vars in dsv4_fp4_mi355x_sglang_mtp.sh
- GPU_MAX_HW_QUEUES=5 placement on the DP-attention branch, confirmed to match the spec-none sibling dsv4_fp4_mi355x_sglang.sh
- The perf-changelog.yaml diff, confirmed to only append at the tail with no existing bytes changed
Extended reasoning...
Overview
This PR touches three files for a single AMD MI355X recipe (dsv4-fp4-mi355x-sglang-mtp): a sglang-rocm image bump in configs/amd-master.yaml, several env-var/flag tweaks in the corresponding launch script (dsv4_fp4_mi355x_sglang_mtp.sh), and an append-only perf-changelog.yaml entry describing the change.
I traced each edit against the PR description and the sibling non-MTP script (dsv4_fp4_mi355x_sglang.sh): the new SHARED_EXPERTS_ARGS array correctly defaults to --enforce-shared-experts-fusion and only switches to --disable-shared-experts-fusion when EP_SIZE>1 (a deliberate divergence from the sibling script, which still disables fusion unconditionally — this matches the PR's stated intent, not a leftover bug). GPU_MAX_HW_QUEUES=5 is set only inside the DP_ATTENTION branch, matching the sibling script's placement as claimed. The two new exported env vars and the new --enable-deepseek-v4-fp4-indexer flag are simple additive changes with no control-flow risk. The perf-changelog.yaml diff adds exactly one entry at the tail; a diff of removed lines confirmed no historical bytes were touched.
Security risks
None. This is a benchmark-recipe/config change with no auth, crypto, or data-handling code involved.
Level of scrutiny
Moderate. The change is mechanical (version bump, flag reorganization, env vars) and follows established patterns in the file, but configs/amd-master.yaml is CODEOWNER-owned (per .github/CODEOWNERS: billishyahao, chunfangamd, seungrokj, yctseng0211, 1am9trash) and the repo's own CONTRIBUTING.md mandates a CODEOWNER sign-off with a specific checklist before merge — the PR author here is one of that file's own codeowners, so an independent codeowner/human look matters for this specific PR regardless of the automated review outcome.
Other factors
I was unable to reach the network from this sandboxed environment to independently confirm the new image tag (lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260907) actually exists on Docker Hub, per the known image-tag pitfall documented in KLAUD_DEBUG.md section 6. No other recipe in amd-master.yaml currently references this exact tag, though the version/date progression (v0.5.18-...-20260902 is the latest pre-existing tag, this one is dated 20260907, five days later and two days before the PR was opened 2026-09-09) is consistent with a legitimate new nightly rather than an invented one. I could not conclusively verify this either way, which is an additional reason a human look is warranted.
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as outdated.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34336290295 |
|
|
/reuse-sweep-run 34336290295 |
|
/reuse-sweep-run |
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
- 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; 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 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.
Additional detail section:
- insert any additional info here
cookbook is at https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4#hw=mi355x&variant=pro&quant=fp4&strategy=high-throughput&nodes=single
Signed: seungrokj
❌❌❌ REJECTED ❌❌❌@seungrokj Two blockers: (1) this PR updates the ✅ Check 0 (CODEOWNER): PASS — @seungrokj is a listed owner of |
4 participants
cc @chunfangamd, @HaiShaw
Summary:
lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260907.--disable-shared-experts-fusionwith--enforce-shared-experts-fusion, and move the disable to the EP > 1 branch only.TORCH_BLAS_PREFER_HIPBLASLT=1andSGLANG_OPT_USE_AITER_BATCHED_GEMM=1.GPU_MAX_HW_QUEUES=5on the DP-attention path, matching the spec-none siblingdsv4_fp4_mi355x_sglang.sh.--enable-deepseek-v4-fp4-indexer.Note
Low Risk
Benchmark launch script, image pin, and search-space tweaks only; no application auth or data-path changes.
Overview
Updates the DSv4 FP4 MI355X SGLang MTP 8K/1K benchmark to the
v0.5.19-rocm720-mi35x-20260907image and refreshes launch flags/env to match the non-MTP sibling recipe.dsv4_fp4_mi355x_sglang_mtp.shnow enables shared-experts fusion by default (--enforce-shared-experts-fusion), only disabling it whenEP_SIZE > 1, adds hipBLASLt/aiter batched GEMM env toggles, turns on--enable-deepseek-v4-fp4-indexer, and on the DP-attention path setsGPU_MAX_HW_QUEUES=5andSGLANG_PREFILL_DELAYER_MAX_PREFILL_BS_WINDOW_SIZE=1.configs/amd-master.yamlpins the new image fordsv4-fp4-mi355x-sglang-mtpand changes the TP8 no-DP-attention sweep from a continuous range toconc-list: [1, 2, 4, 16, 32](dropping concurrency 8).perf-changelog.yamlrecords these changes for PR #2910.Reviewed by Cursor Bugbot for commit 35daccd. Bugbot is set up for automated code reviews on this repo. Configure here.