AMD AgentX MI355X: GPU sanity dedup, HiCache tuning, node-exclude scoping, and DSv4 disagg image bump - #2643
Conversation
…_gpu_clean Removes helpers/gpu_sanity.sh, whose VRAM-drain check duplicated the wait_for_amd_gpu_clean() gate added to benchmark_lib.sh in #2490. Wires that shared function into job.slurm's pre-flight gate instead, guarding against benchmark_lib.sh's unrelated agentic KV_OFFLOADING validation by unsetting IS_AGENTIC/SCENARIO_TYPE before sourcing it standalone. Co-authored-by: Cursor <cursoragent@cursor.com>
…he image - models.yaml (DeepSeek-V4-Pro-AgentX): move swa-full-tokens-ratio from a global base_flag to per-role prefill/decode fields, add --enable-two-batch-overlap to dp_flags, raise prefill mem_fraction_static to 0.85 and no_dp max_running_requests to 128, enable decode CUDA graphs. - dsv4_fp4_mi355x_sglang-disagg.sh: lower default HICACHE_RATIO to 3 and trim the FORCE_HICACHE_RATIO rationale comment. - amd-master.yaml: bump the dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp image, extend the MTP sweep conc-list to include 64, and comment out the superseded TP8/EP1 single-worker search-space arm. Co-authored-by: Cursor <cursoragent@cursor.com>
…xed global list Known-bad nodes only affect Kimi-K2.5-MXFP4 on vllm-disagg; other framework/model combos no longer get any nodes excluded by default. Mapping lives in node_excludes.yaml; SLURM_EXCLUDE_NODES still overrides it. Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Theresa Shan <theresa.shan@amd.com> bump image to aigmkt/mori-dev:sglang-0.5.17-rocm720-mi35x-mori-0814 Signed-off-by: Theresa Shan <theresa.shan@amd.com> add one more sweep dot Signed-off-by: Theresa Shan <theresa.shan@amd.com> Update amd-master.yaml
|
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 |
1 similar comment
|
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.
Approving — the three findings are all self-contained nits (silent-failure diagnostics gap, a sensitivity regression with an existing escape hatch, and a dead config key) rather than functional blockers, and the PR is benchmark/tuning infra rather than production-serving code.
Extended reasoning...
Overview
This PR touches AMD multi-node benchmark tooling only: it wires the shared wait_for_amd_gpu_clean helper into job.slurm's pre-flight gate (replacing the standalone gpu_sanity.sh), tunes DeepSeek-V4-Pro-AgentX HiCache/CUDA-graph settings in models.yaml, scopes submit.sh's node exclusions to specific (framework, model) combos via a new node_excludes.yaml, and bumps an image tag plus extends a concurrency sweep in configs/amd-master.yaml. No production inference-serving code, auth, or data-handling paths are touched.
Security risks
None. This is CI/benchmark orchestration bash and YAML config running on internal SLURM/Docker infrastructure the author controls; there's no user-facing input, auth, or data exposure surface. The python3 -c node_excludes lookup interpolates values from the same repo's YAML/env vars, not external/untrusted input.
Level of scrutiny
Medium — this is benchmark tooling (not production-critical), but it does touch a safety gate (GPU drain check) and a known-bad-node exclusion mechanism, so correctness matters for CI reliability even though a bad merge here only wastes benchmark runs rather than affecting production traffic. The bug-hunting pass found three issues, all explicitly scoped as nits: a silent-failure mode if python3/PyYAML is missing on the bare submit host (mitigated by an existing identical pattern elsewhere in the repo), a reduction in the GPU-drain gate's absolute-GB sensitivity when switched to the shared percentage-based helper (mitigated by the existing SKIP_GPU_SANITY escape hatch and the fact it only misses a narrow 8-28GB band), and a dead/unwired decode.disable_cuda_graph config key that is a no-op today since its value is false (matching current always-on behavior).
Other factors
None of the three findings represent a functional regression that would break CI or produce incorrect benchmark results today; they're robustness/precision gaps worth fixing but not blocking. The PR is otherwise a straightforward consolidation (removing a duplicate GPU check) and parameter tuning change following established patterns in this file.
…quests to 64 Restores the pre-2fd1ce507 value; 128 was too aggressive for the no_dp prefill role. Co-authored-by: Cursor <cursoragent@cursor.com>
… trusted Checks python3/PyYAML availability upfront and the exit status of the python3 invocation, falling back to an awk parser (mirroring job.slurm's style) when python3/PyYAML are unavailable. Previously a missing python3 or yaml module would silently yield an empty exclude list, reintroducing the known-bad-node scheduling issue with zero diagnostic. Co-authored-by: Cursor <cursoragent@cursor.com>
…ek-V4-Pro-AgentX decode Decode CUDA graphs are enabled by default; the explicit false was a no-op alongside the existing dp/no_dp cuda_graph_bs_range settings. Co-authored-by: Cursor <cursoragent@cursor.com>
ichbinblau
added
AMD
agentx
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32269076444 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32269076444 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32324106619 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32269076444 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32269076444 |
|
/reuse-sweep-run 32269076444 |
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/32269076444
- Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. https://github.com/semiAnalysisAI/InferenceX/actions/runs/32269076444
- 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
Signed: billishyahao
❌❌❌ REJECTED ❌❌❌@billishyahao — the sign-off uses a stale copy of the checklist: it is missing the ✅ Check 0 (CODEOWNER): PASS — |
billishyahao
left a comment
billishyahao
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. https://github.com/semiAnalysisAI/InferenceX/actions/runs/32269076444
- Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. https://github.com/semiAnalysisAI/InferenceX/actions/runs/32269076444
- 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
Signed: billishyahao
✅✅✅ Verdict: PASS ✅✅✅The sign-off has been re-posted from the current checklist template (the previously missing ✅ Check 0 (CODEOWNER): PASS — |
|
/stage-results 32269076444 |
|
@cquil11 |
|
/stage-results 32269076444 |
|
@cquil11 staged run 32269076444: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-19~r32269076444 This run remains available across future |
|
/reuse-sweep-run 32269076444 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32444234325 |
Keep only the image bump, added concurrency points, and performance tuning deltas. 中文：精简 PR 2643 的性能变更日志，仅保留镜像更新、新增并发点和性能参数调整。
5 participants
Summary
wait_for_amd_gpu_clean()intojob.slurm's pre-flight gate and removes the now-duplicatehelpers/gpu_sanity.shVRAM-drain check.swa-full-tokens-ratio,--enable-two-batch-overlap, prefillmem_fraction_static, decode CUDA graphs) and lowers the defaultHICACHE_RATIO.submit.shnode exclusions to the (framework, model) combos that actually need them (Kimi-K2.5-MXFP4 on vllm-disagg) vianode_excludes.yaml, instead of excluding known-bad nodes globally.dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtpSGLang image and merges the duplicate spec-none/MTP config entries inamd-master.yamlinto one, extending the MTP concurrency sweep.Test plan
job.slurmpre-flight GPU-drain gate still triggers correctly withhelpers/gpu_sanity.shremovedsubmit.shonly excludes nodes for Kimi-K2.5-MXFP4 + vllm-disagg, andSLURM_EXCLUDE_NODESstill overridesnode_excludes.yamldsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtpe2e sweep to validate the merged config and new imageMade with Cursor