glm5.1-fp4-mi355x-sglang: bump image to v0.5.13.post1-20260622 + enable aiter allreduce fusion#1905
glm5.1-fp4-mi355x-sglang: bump image to v0.5.13.post1-20260622 + enable aiter allreduce fusion#1905jiacao-amd wants to merge 3 commits into
Conversation
…le aiter allreduce fusion - Bump sglang image v0.5.12.post1-20260529 -> v0.5.13.post1-20260622: the older image's aiter (dev98) does not tolerate the host ROCm 7.1.1 x container 7.2.0 mismatch on MI355X and produces garbled output; the newer aiter (dev240) does. The newer build is also faster at decode. - Add --enable-aiter-allreduce-fusion: fuse TP all-reduce + RMSNorm (atom-style local_device_load_rmsnorm). Marginal perf; gated by CI GSM8K eval. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
jiacao-amd
requested review from
1am9trash,
billishyahao,
chunfangamd,
seungrokj and
yctseng0211
as code owners
|
Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you
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. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers. If additional help is needed, PR authors can reach out to core maintainers over Slack. 感谢你的贡献！对于 vLLM 与 SGLang，请确保你的 recipe 与官方 vLLM recipes 和/或 SGLang cookbook 保持一致 如果不一致，请先创建一个 PR，之后我们才能将你的单节点 PR 合并到 master 分支。让我们确保文档保持一流水准，使整个 ML 社区都能从你的辛勤工作中受益！谢谢
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。如果选择重新运行失败的任务，PR 作者有责任确保其最终通过。参见 GitHub 关于重新运行失败任务的文档：https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow 一般而言，PR 作者应先向相应公司的 CODEOWNERS 请求审阅并获得 PR 批准，然后再请求核心维护者审阅。 如需更多帮助，PR 作者可通过 Slack 联系核心维护者。 |
|
/sweep test-config --config-files .github/configs/amd-master.yaml --config-keys glm5.1-fp4-mi355x-sglang |
|
@jiacao-amd Kicking off a sweep. Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28049943426 |
Perf: image bump (v0.5.12→v0.5.13) +
|
| tp | conc | tput/gpu old→new | Δ tput | intvty old→new | Δ intvty |
|---|---|---|---|---|---|
| 2 | 4 | 186→201 | +7.8% | 48.3→53.3 | +10.2% |
| 2 | 8 | 312→346 | +10.9% | 40.3→44.8 | +11.0% |
| 2 | 16 | 498→539 | +8.0% | 32.0→34.6 | +8.1% |
| 2 | 32 | 760→818 | +7.6% | 24.6→26.5 | +7.6% |
| 2 | 64 | 1134→1195 | +5.4% | 18.3→19.3 | +5.4% |
| 2 | 128 | 1630→1712 | +5.0% | 13.2→13.8 | +5.1% |
| 2 | 256 | 2077→2285 | +10.0% | 8.4→9.2 | +10.1% |
| 4 | 4 | 109→122 | +11.4% | 56.7→63.2 | +11.5% |
| 4 | 8 | 195→215 | +10.2% | 50.5→55.7 | +10.4% |
| 4 | 16 | 332→349 | +4.9% | 42.6→44.7 | +4.9% |
8k / 1k
| tp | conc | tput/gpu old→new | Δ tput | intvty old→new | Δ intvty |
|---|---|---|---|---|---|
| 2 | 4 | 705→764 | +8.5% | 41.7→45.5 | +9.0% |
| 2 | 8 | 1121→1219 | +8.7% | 33.3→36.3 | +9.0% |
| 2 | 16 | 1619→1727 | +6.7% | 23.9→25.5 | +6.7% |
| 2 | 32 | 2213→2362 | +6.7% | 16.4→17.5 | +6.8% |
| 2 | 64 | 2878→3050 | +6.0% | 10.5→11.1 | +6.0% |
| 2 | 128 | 3187→3384 | +6.2% | 8.1→8.6 | +6.1% |
| 2 | 256 | 3034→3397 | +12.0% | 7.5→8.5 | +12.3% |
| 4 | 4 | 423→464 | +9.8% | 49.8→54.7 | +9.9% |
| 4 | 8 | 656→781 | +19.2% | 38.7→46.2 | +19.2% |
| 4 | 16 | 1082→1169 | +8.0% | 31.9→34.4 | +7.8% |
Average throughput/GPU gain: +8.1% (1k/1k), +9.2% (8k/1k) — consistent +5–19% across all concurrencies. (Most of the gain is the newer aiter in v0.5.13; AR fusion contributes marginally.)
|
/sweep test-config --config-files .github/configs/amd-master.yaml --config-keys glm5.1-fp4-mi355x-sglang |
|
@jiacao-amd Kicking off a sweep. Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28058547561 |
|
@jiacao-amd use the pr label of "full-sweep-enabled" instead of /sweep plz and then ping an AMD CODEOWNER for inital review https://github.com/SemiAnalysisAI/InferenceX/blob/main/docs/PR_REVIEW_CHECKLIST.md |
done |
jiacao-amd
force-pushed
the
glm5.1-fp4-mi355x-image-bump-arfusion
branch
from
f492841 to
65a9952
Compare
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28083001624 |
5 similar comments
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28083001624 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28083001624 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28083001624 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28083001624 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28083001624 |
3 participants
What
glm5.1-fp4-mi355x-sglangimagev0.5.12.post1-20260529->v0.5.13.post1-20260622.--enable-aiter-allreduce-fusionto the GLM5.1 fp4 mi355x launch script.Why
--enable-aiter-allreduce-fusionfuses TP all-reduce + RMSNorm (atom-stylelocal_device_load_rmsnorm).Validation
Trigger CI (incl. GSM8K eval) on this PR: