perf(amd): add DP-attention band to DSV4 ATOM MI355X AgentX MTP - #2778
Conversation
Split the DeepSeek-V4-Pro FP4 ATOM AgentX MTP sweep into two attention bands per the ROCm/ATOM recipe: tensor-parallel attention for concurrency 1-16, and DP attention (--enable-dp-attention --enable-tbo plus the ATOM_DP_* routing knobs and AIPerf session-affinity vars) for concurrency 48-256. The mode is selected by the config's dp-attn, propagated as DP_ATTENTION, so the recipe drops its DP_ATTENTION=false guard. Bump the ATOM image to nightly_202608280858. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
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 |
2 similar comments
|
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 |
|
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 |
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
seungrokj
added
AMD
full-sweep-fail-fast
agentx
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
| # golden_al_distribution/dsv4_mtp.yaml: thinking_on, 3 draft tokens -> AL 2.49 | ||
| # --spec-decode-acceptance-length 2.49. | ||
| # https://github.com/ROCm/ATOM/pull/1948 | ||
| # https://github.com/SemiAnalysisAI/InferenceX/blob/main/golden_al_distribution/dsv4_mtp.yaml |
There was a problem hiding this comment.
High-conc DP warmup grace missing
Medium Severity
The new DP-attention band runs concurrency 48 through 256 with the default 1800s warmup grace. This same config previously documented that AgentX saturation at concurrency 32 and 48 needs AGENTIC_WARMUP_GRACE_PERIOD=3600 so long-prefix warmup can drain, and the MI355X DSv4 SGLang sibling already applies that override at CONC >= 32. The new high-concurrency DP arms can hit warmup timeout and fail before profiling.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 8757b3e. Configure here.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Additional findings (no inline location):
-
🔴
perf-changelog.yaml— Recipe change (image bump, guard removal, new DP-attention search-space row, removed AIPERF threshold vars) has no perf-changelog.yaml entry, violating AGENTS.md's mandatory append-only changelog rule for every recipe addition/modification. [also at: configs/amd-master.yaml:1350 - No perf-changelog.yaml entry added for a benchmark-affecting recipe change (image bump + search-space split + new…]Extended reasoning...
AGENTS.md requires every recipe-affecting change to append a perf-changelog.yaml entry; this PR touches both the benchmark script and configs/amd-master.yaml but adds none, so downstream consumers of the changelog lose the record of why the DSV4 ATOM MTP recipe was split into TP/DP bands and the image was bumped.
Verification: normal (convention violation of a stated non-negotiable invariant). AGENTS.md line 22 states: "Every change that can affect benchmark performance and every recipe addition or modification requires a new
perf-changelog.yamlentry. The file is append-only and byte-sensitive." This is listed under "Non-negotiable benchmark invariants." The diff (`git diff… | normal (convention violation the…
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33165527762 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33165527762 |
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…MI355X MTP Bump dataset/profile-configure timeouts to 1800s and cancel-drain to its 300s ceiling for long agentic cold-cache runs, and enable realtime metrics. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
uvicorn's 5s idle keep-alive closes AIPerf's pooled sockets during long-backlog warmup; the reused write hits ECONNRESET (104) and one such reset on a root AgentX warmup request aborts the entire run. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.
There are 3 total unresolved issues (including 2 from previous reviews).
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 4c28e5d. Configure here.
Re-enable the TP c1-16 and DP-attention c48-256 bands after the keep-alive fix, reverting the temporary dp-attn c256-only e2e scoping. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33376779864 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33376779864 |
|
/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. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33376779864
- Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33376779864/job/99523227952
- 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
Recipe: https://github.com/ROCm/ATOM/blob/main/recipes/DeepSeek-V4-Agentic-InferenceX.md
Signed: @chunfangamd
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — @chunfangamd is a listed owner of |
4 participants
Summary
DeepSeek-V4-Agentic-InferenceX.mdrecipe:dp-attn: false) for concurrency1, 2, 4, 8, 16dp-attn: true) for concurrency48, 64, 96, 128, 256, adding--enable-dp-attention --enable-tboplus theATOM_DP_*routing knobs (ATOM_DP_SESSION_AFFINITY,ATOM_DP_LB_REQ_EQUIV,ATOM_ENABLE_PREFILL_DELAYER,ATOM_PREFILL_DECODE_INTERVAL), the two-batch-overlap prerequisites (GPU_MAX_HW_QUEUES,ATOM_NUMA_BIND), and the AIPerf session-affinity vars.dp-attn, propagated to the script asDP_ATTENTION; the recipe's oldDP_ATTENTION=falsehard-fail guard is removed.rocm/atom-dev:nightly_202608201032→nightly_202608280858.Test plan
🤖 Generated with Claude Code
Note
Low Risk
Benchmark config and launch-script changes only; DP session affinity is correctness-critical for agentic results but does not affect production serving paths.
Overview
Splits the dsv4-fp4-mi355x-atom-agentic-mtp AgentX sweep into tensor-parallel attention (
dp-attn: false, concurrency 1–16) and DP attention (dp-attn: true, 48–256), replacing the old single TP-only list through conc 48. Bumps the ATOM image torocm/atom-dev:nightly_202608280858.The recipe script stops failing when
DP_ATTENTIONis true and branches on config-drivenDP_ATTENTION: the DP band turns on--enable-dp-attention/--enable-tbo, ATOM DPA routing env vars (including mandatoryATOM_DP_SESSION_AFFINITY), and AIPerf correlation-id session headers so multi-turn prefix KV stays on one rank. It also lengthens server--timeout-keep-aliveand refreshes several AIPerf client timeouts/settings for long AgentX stalls; some prior failure/warmup AIPerf thresholds are dropped.perf-changelog.yamldocuments the matrix and recipe changes.Reviewed by Cursor Bugbot for commit b041df8. Bugbot is set up for automated code reviews on this repo. Configure here.