[AMD][MI35X] Add HiCache and TP2/EP1 arms to the Qwen3.5 MXFP4 MI355X AgentX sweep - #2693
[AMD][MI35X] Add HiCache and TP2/EP1 arms to the Qwen3.5 MXFP4 MI355X AgentX sweep#2693yichiche wants to merge 6 commits into
Conversation
… AgentX sweep Add a HiCache host-DRAM KV tier to the AgentX script, gated on KV_OFFLOADING=dram with KV_OFFLOAD_BACKEND=hicache. The ratio 1.5 / write_through / direct / page_first_direct combination is the one already validated on cluster:mi355x-amds by the dsv4 and glm5.2 AgentX recipes, and every knob stays overridable. The arm keeps --page-size 16 so only the cache tier differs between the kvnone and kvdram points. Replace the TP2/EP2 arm with TP2/EP1 to match qwen3.5-fp4-b200-sglang-agentic-mtp's 2-GPU point, and add kvdram arms at TP4 concurrency 40-64 and TP2 concurrency 20-32. Each kvdram arm repeats its kvnone neighbour's last concurrency so the host tier's gain is measured at a fixed operating point.
|
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.
Automated review found no functional bugs — the only finding is a nit-level pr-link metadata mistake in the changelog entry (already flagged inline). Given a candidate accuracy-eval-arm-selection concern was raised and ruled out this run, leaving this note for the record.
Extended reasoning...
Beyond the inline nit, I also checked whether the accuracy eval row silently moved to a different arm (kvnone to the new HiCache conc=64 arm) in configs/amd-master.yaml — the eval row selection is independent of which arm this PR adds, so it's not affected.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32337711416 |
4 similar comments
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32337711416 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32337711416 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32337711416 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32337711416 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32456812122 |
|
/reuse-sweep-run |
|
@1am9trash sgl-project/sglang#36245 merged. |
1am9trash
left a comment
1am9trash
left a comment
There was a problem hiding this comment.
As a PR reviewer and CODEOWNER, I have reviewed this and have:
- x ] 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. Link: https://inferencex.semianalysis.com/inference?unofficialRun=32456812122
- Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. Link: https://inferencex.semianalysis.com/evaluation?unofficialRun=32456812122
- 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:
- Recipe PR (merged): sgl-project/sglang#36245
Signed: @1am9trash
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — @1am9trash is a listed owner of |
|
Hi, @cquil11 |
4 participants
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32456812122
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32456812122
Motivation
Two structural gaps remain between
qwen3.5-fp4-mi355x-sglang-agentic-mtpand its B200 counterpartqwen3.5-fp4-b200-sglang-agentic-mtp, and both make the current AgentX comparison misleading.First, KV offload. The B200 arm carries a HiCache host-DRAM KV tier and rides it past the GPU-KV ceiling to concurrency 64+, while MI355X runs GPU-resident KV only and stops at concurrency 40. The MI355X curve therefore ends where the B200 curve is still climbing, and the difference measured at the top of the sweep is a cache-tier difference rather than a hardware difference.
Second, the 2-GPU point. B200 sweeps TP2/EP1; MI355X sweeps TP2/EP2. Those are not the same partitioning, so the 2-GPU columns are not comparable across the two fleets.
Modifications
Add a HiCache host-DRAM KV tier to
benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh, gated onKV_OFFLOADING=dramwithKV_OFFLOAD_BACKEND=hicachethrough the sharedrequire_agentic_kv_offload_backendhelper.KV_OFFLOADINGandTOTAL_CPU_DRAM_GBjoin the script'scheck_env_varslist, matching the B200 and dsv4 AgentX recipes.The tier uses
--enable-hierarchical-cachewith ratio 1.5,write_through,directio, andpage_first_directlayout. That combination is the one already validated oncluster:mi355x-amdsbydsv4_fp4_mi355x_sglang_mtp.shandglm5.2_fp4_mi355x_sglang_mtp.sh; every knob stays overridable via environment variable for tuning. Unlike the B200 sibling this arm keeps--page-size 16, so the kvnone and kvdram points differ only in the cache tier.In
configs/amd-master.yaml, replace the TP2/EP2 arm with TP2/EP1 and add two kvdram arms:Each kvdram arm repeats its kvnone neighbour's last concurrency (40 at TP4, 20 at TP2) so the host tier's gain is measured at a fixed operating point rather than inferred across a concurrency step. This follows the same idiom as
dsv4-fp4-mi355x-sglang-agentic-mtp, which repeats concurrency 16 across its two arms.The sweep grows from 16 to 24 points.
Accuracy Tests
No accuracy-affecting logic changes in this repo. The kvdram arms change where KV lives, not what is computed; the AgentX eval rows continue to run real target-model verification.
Benchmarking
Repo validation was run locally:
python -m pytest utils/matrix_logic/ -q→ 232 passed.bash -n benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh→ clean.python utils/matrix_logic/generate_sweep_configs.py full-sweep --config-files configs/amd-master.yaml --model-prefix qwen3.5 --precision fp4 --scenario-type agentic-coding→ 24 configs: TP4/EP1 kvnone 10, TP2/EP1 kvnone 6, TP4/EP1 kvdram-hicache 4, TP2/EP1 kvdram-hicache 4. No TP2/EP2 points remain.End-to-end MI355X AgentX numbers will come from the sweep triggered on this PR (
full-sweep-fail-fast). HiCache ratio 1.5 is the starting point inherited from the sibling recipes on this cluster; if the host pool is oversubscribed or the hit rate is poor,HICACHE_RATIOis the first knob to retune.Note
Low Risk
Benchmark and sweep configuration only; HiCache wiring follows existing AgentX scripts with no changes to model accuracy or serving auth paths.
Overview
Aligns the Qwen3.5 MXFP4 MI355X AgentX benchmark with the B200 recipe so cross-fleet comparisons are fair.
The agentic launch script now requires
KV_OFFLOADINGandTOTAL_CPU_DRAM_GB, and when the sweep selects HiCache it appends SGLang hierarchical-cache flags (ratio 1.5, write-through, direct I/O, page-first layout) using the samerequire_agentic_kv_offload_backendpattern as other AgentX scripts.In
configs/amd-master.yaml, the TP2/EP2 GPU-only arm is replaced with TP2/EP1, and two new host-DRAM KV (hicache) arms extend concurrency beyond the GPU-KV ceiling (TP4 at 40–64, TP2 at 20–32).perf-changelog.yamlrecords the sweep change (16 → 24 points).Reviewed by Cursor Bugbot for commit 2b669eb. Bugbot is set up for automated code reviews on this repo. Configure here.