Refresh the DSv4 MI355X SGLang AgentX arm to v0.5.18 and split TP4/TP8 by concurrency

[karverma-amd]

Summary

Refreshes dsv4-fp4-mi355x-sglang-agentic-mtp onto the v0.5.18 MI355X image with the serving flags validated on the AMD side, and splits the search space so TP4 covers the low-concurrency end instead of TP8 duplicating it.

Replaces #2710, which was opened from a fork.

Image and serving flags

• lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260813 → lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260822
• Enable SGLANG_OPT_USE_AITER_BATCHED_GEMM=1, the aiter batched GEMM for the absorbed MLA projections, which defaults off in environ.py
• mem-fraction-static 0.85 → 0.89, enlarging the full-attention KV pool. The DSv4 compressor state pools are sized from that pool and allocated after it, outside this budget, so the remainder still has to cover them — which is why this stops short of 0.90.
• --disable-shared-experts-fusion → --enforce-shared-experts-fusion. The tuned MoE tables in this image cover the resulting (inter_dim 384, expert 385, topk 7) shape at the token sizes this workload hits.

Search space

• Add a TP4 arm at concurrency [1, 2, 4, 8, 10] with no host KV tier, where halving the GPUs per replica raises throughput per GPU
• Reduce the TP8 no-offload arm from [1, 2, 4, 8, 16] to [16], so TP8 picks up where TP4 leaves off
• Net effect on published data: drops TP8 concurrency 1, 2, 4 and 8; adds TP4 concurrency 1, 2, 4, 8 and 10. The hicache arm is unchanged at [16, 32, 48], and concurrency 16 still appears on both TP8 arms to isolate the host KV tier's gain.

TP4 doubles per-rank weights and leaves far less room for KV, which is why it stops at concurrency 10 and ships no hicache variant.

runner (cluster:mi355x-amds) is unchanged and no directory structure changed. The key's comment is also corrected: it no longer claims the image matches the disagg AgentX entry (that key is still on v0.5.17-...-20260817), and "pure TP8 only" becomes "tensor-parallel only", which is what it meant — the exclusion is DP-attention, not TP4.

Test plan

• bash -n benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh
• configs/amd-master.yaml and perf-changelog.yaml both parse
• Config generation yields 9 cells on the new image: TP4 at conc 1/2/4/8/10 no-offload, TP8 at conc 16 no-offload plus 16/32/48 hicache
• process_changelog.py --base-ref main --head-ref HEAD triggers all 9 cells plus the conc-48 agentic eval
• Sweep run on MI355X — this image and flag set has not yet been run on this key

Made with Cursor

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32688409514
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32688409514

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f49b272. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32724927772
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32724927772

[1am9trash]

/reuse-sweep-run 32724927772

[1am9trash]

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. Link: https://inferencex.semianalysis.com/inference?unofficialRun=32724927772
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. Link: https://inferencex.semianalysis.com/evaluation?unofficialRun=32724927772
• 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 at https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4#hw=mi355x&variant=pro&quant=fp4&strategy=high-throughput&nodes=single

Signed: @1am9trash

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — configs/amd-master.yaml is specifically owned by @1am9trash; the other changed paths carry only the catch-all, which a recognized CODEOWNER satisfies.
✅ Check 1 (passing sweep on in-PR commit): PASS — run 32724927772 on head 58bbb80 executed all 8 agentic / benchmark cells plus agentic eval / with conclusion success (agentic-only PR, so the single-node */ / eval / lanes were legitimately not selected).
✅ Check 2 (evals pass): PASS — gsm8k em_strict 0.9674 (n_eff 1319) ≥ the 0.91 dsv4 threshold, run on this PR's image lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260822.
✅ Check 3 (recipe merged & args match): PASS — the linked published SGLang cookbook page documents MI355X Pro FP4 single-node with the same major args: --attention-backend dsv4, --page-size 256, --kv-cache-dtype fp8_e4m3, --enforce-shared-experts-fusion, EAGLE 3-1-4, SGLANG_OPT_USE_AITER_BATCHED_GEMM; the TP8 arm matches the verified TP-only cell, TP4 is a documented parallelism option for this variant, and the hicache arm matches the cookbook's AMD hicache spec (page_first_direct + direct). Informational InferenceX tuning diffs only: mem-fraction 0.89 vs 0.90, swa-full-tokens-ratio 0.10 vs 0.15, hicache ratio 1.5, image tag v0.5.18 vs the page's v0.5.17.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run 32724927772 posted by @1am9trash (COLLABORATOR).
✅ Check 5 (latest checklist): PASS — every item of the current docs/PR_REVIEW_CHECKLIST.md template is present and checked.
✅ Check 6 (upstream image / engine-first): PASS — lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260822 is the upstream lmsysorg org, and the entry is framework: sglang so engine-first ordering is inherently satisfied.
✅ Check 7 (deprecated models): PASS — dsv4 agentic-coding MTP is the active, published arm per MODELS.md as of 2026-08-24.
✅ Check 8 (no architecture hacks): PASS — no --hf-overrides or model-config edits; --enforce-shared-experts-fusion fuses the same computation without removing FLOPs.
✅ Check 9 (chat templates): PASS — AgentX trace replay drives chat-completions with the model's native chat template (no truncating template override).
✅ Check 10 (no engine patches): PASS — no patching of the serving stack; the script installs only client-side harness deps.
✅ Check 11 (golden AL): PASS — throughput runs pin SGLANG_SIMULATE_ACC_LEN=2.49 (match-expected, real-draft-token), equal to golden_al_distribution/dsv4_mtp.yaml thinking_on / 3 speculative steps; eval-only runs keep real target verification.
➖ Check 12 (append-only): N/A — the new perf-changelog entry does not use append-only: true.

[1am9trash]

Hi, @functionstackx
May you help to review this? Thanks.

[functionstackx]

@1am9trash plz ping @cquil11 or @adibarra for review plz

[cquil11]

/stage-results 32724927772

[github-actions]

@cquil11 staged run 32724927772: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-08-24~r32724927772

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow