[AMD][MI35X] Bump Qwen3.5 MXFP4 MI355X SGLang AgentX image to v0.5.17-20260818

[yichiche]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32253800160
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32253800160

Motivation

qwen3.5-fp4-mi355x-sglang-agentic-mtp is the AgentX (agentic-coding) arm for Qwen3.5-397B-A17B-MXFP4 on MI355X, and it is still pinned to lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260811. The fixed-seq-len arms for the same model are moving to v0.5.17-rocm720-mi35x-20260818 in #2640, so the agentic arm is now the only MI355X Qwen3.5 MXFP4 SGLang submission left on an older nightly. This bumps it to the same image so the agentic numbers are measured against the same upstream SGLang + aiter build as the fixed-seq-len ones.

20260818 is the newest rocm720-mi35x tag published on Docker Hub at the time of writing.

Modifications

Bump the image: value of qwen3.5-fp4-mi355x-sglang-agentic-mtp in configs/amd-master.yaml from lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260811 to lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260818. Nothing else in the config changes — model, runner: cluster:mi355x-amds, dram-utilization, TP/EP search space, kv-offloading: none, and the TP2/TP4 concurrency lists are untouched.

Also add --kv-cache-dtype fp8_e4m3 to benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh, matching the fp8 KV cache that #2640 applies to the fixed-seq-len arms. The remaining MTP tunings from that PR are not carried over: --mamba-ssm-dtype bfloat16 (already exported as SGLANG_MAMBA_SSM_DTYPE), --max-prefill-tokens / --chunked-prefill-size 32768 and --tokenizer-worker-num 6 are already present in the agentic script, --context-length ISL+OSL+20 does not apply to variable-length trace replay, and --scheduler-recv-interval / --stream-interval are deliberately set to 30 / 50 here because AgentX replay is long-running with many concurrent streams.

Append the corresponding perf-changelog.yaml trigger at the end of the file with scenario-type: agentic-coding, so the sweep dispatches this entry through sweep-agentic rather than the fixed-seq-len path.

This PR is deliberately kept separate from #2640: the agentic arm runs on the cluster:mi355x-amds runner pool rather than mi355x, and under full-sweep-fail-fast a failure on one arm would take the other arm's results down with it.

Accuracy Tests

No accuracy-affecting logic changes in this repo — the change is limited to the container tag and the changelog trigger. Model correctness is covered by the upstream SGLang release validation for the v0.5.17 ROCm image.

Benchmarking

Repo validation was run locally:

• python -m pytest utils/matrix_logic/ -q → 232 passed.
• 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 emits 16 agentic configs (6 on TP2/EP2, 10 on TP4/EP1), all referencing lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260818 on cluster:mi355x-amds.
• bash -n benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh → clean.

End-to-end MI355X AgentX numbers will come from the sweep triggered on this PR (full-sweep-fail-fast).

[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

[claude]

Straightforward image tag bump with one minor changelog nit flagged inline (wrong pr-link, copy-pasted from the preceding entry) — not functionally significant since scenario-type/config-keys drive the actual sweep dispatch.

Extended reasoning...

Overview

This PR bumps the image: tag for qwen3.5-fp4-mi355x-sglang-agentic-mtp in configs/amd-master.yaml from the 20260811 to the 20260818 SGLang ROCm nightly, and appends a matching trigger entry to perf-changelog.yaml. No search-space, runner, or scenario-type changes — verified against the diff that only the image string changed in the config, and the changelog addition is a pure append at EOF.

Security risks

None. This is a container image tag change and a changelog metadata append; no code paths, auth, or secrets are touched.

Level of scrutiny

Low. This is a mechanical, single-value config bump of the same kind routinely made across this file (see the many sibling image-bump entries), consistent with the PR's own stated motivation (aligning the agentic arm with the fixed-seq-len arms' image from #2640). The one issue found — the new changelog entry's pr-link pointing to #2576 instead of #2677, a copy-paste leftover from the immediately preceding entry — is a traceability nit that doesn't affect sweep dispatch (config-keys/scenario-type drive that), so it doesn't block approval.

Other factors

The PR author's own validation (pytest suite green, generate_sweep_configs dry run showing 16 correctly-routed agentic configs referencing the new image) covers the mechanical correctness of the change. No outstanding review comments to address.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32253800160
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32253800160

[yichiche]

Cookbook: sgl-project/sglang#35445

[1am9trash]

/reuse-sweep-run 32253800160

[1am9trash]

Reuse 32253800160 run perf.
Only solving conflict issue after it.

[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. https://inferencex.semianalysis.com/inference?unofficialRun=32253800160
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. https://inferencex.semianalysis.com/evaluation?unofficialRun=32253800160
• 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:

Cookbook PR: sgl-project/sglang#35445

Signed: @1am9trash

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — @1am9trash is a listed owner of configs/amd-master.yaml; remaining paths are catch-all only.
✅ Check 1 (sweep on in-PR commit): PASS — commit 1e71321 is in this PR with all 16 agentic / jobs and the agentic eval / job success in run 32253800160 (single-node *//eval / lanes are empty-matrix skips for this agentic-only PR).
✅ Check 2 (evals pass): PASS — GSM8K em_strict 0.976 (n_eff 1319) on TP4 c40 MTP, run on lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260818, the exact image this PR pins.
✅ Check 3 (recipe merged + args match): PASS — sgl-project/sglang#35445 MERGED 2026-08-19; cookbook documents MI355X MXFP4 with --attention-backend aiter, --page-size 16, --kv-cache-dtype fp8_e4m3, and identical MTP draft config (steps 3 / topk 1 / draft-tokens 4; NEXTN is the same native-MTP path as EAGLE). Concurrency caps, radix-cache handling, and the TP4 sweep lane are InferenceX sweep tuning (informational only).
✅ Check 4 (reuse command): PASS — /reuse-sweep-run 32253800160 posted by @1am9trash (COLLABORATOR).
✅ Check 5 (latest checklist): PASS — every current-template item present and checked.
✅ Check 6 (upstream image / engine-first): PASS — framework: sglang with upstream lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260818; engine-first ordering not applicable to an SGLang entry.
✅ Check 7 (deprecated models): PASS — qwen3.5 fp4 agentic-coding is active per MODELS.md as of 2026-08-20.
✅ Check 8 (no architecture hacks): PASS — only change is --kv-cache-dtype fp8_e4m3 (lower precision, evals pass); no FLOPs-reducing overrides.
✅ Check 9 (chat template): PASS — replay appends --apply-chat-template.
✅ Check 10 (no engine patches): PASS — no patching; diff adds one server flag, an image tag bump, and a changelog entry.
✅ Check 11 (golden AL): PASS — SGLANG_SIMULATE_ACC_LEN=3.39 with match-expected/real-draft-token, matching golden qwen3.5_mtp.yaml thinking_on @ 3 speculative tokens = 3.39; eval path correctly keeps real verification.
➖ Check 12 (append-only): N/A — the new perf-changelog entry does not use append-only: true.

[functionstackx]

/reuse-sweep-run

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — @1am9trash is a listed owner of configs/amd-master.yaml; the other changed paths carry only the catch-all, which a recognized CODEOWNER satisfies.
✅ Check 1 (sweep on in-PR commit): PASS — commit 1e71321 (still in this PR) has all 16 executed agentic / benchmark check-runs plus agentic eval / green in run 32253800160.
✅ Check 2 (evals pass): PASS — GSM8K em_strict 0.9757 ≥ the 0.94 qwen3.5 bar (utils/evals/thresholds.yaml), run on this PR's image lmsysorg/sglang-rocm:v0.5.17-rocm720-mi35x-20260818.
✅ Check 3 (recipe): PASS — sgl-project/sglang#35445 MERGED 2026-08-19; the published Qwen3.5 MI355X MXFP4 cookbook recipe matches the major args (aiter attention backend, --kv-cache-dtype fp8_e4m3, MTP steps 3 / topk 1 / draft 4 — recipe's NEXTN is the SGLang alias for this EAGLE MTP path, AITER env set). TP/EP sweep lanes and scheduler/stream/radix knobs are InferenceX sweep tuning (informational only).
✅ Check 4 (reuse command): PASS — /reuse-sweep-run 32253800160 posted by @1am9trash (COLLABORATOR) and /reuse-sweep-run by @functionstackx (MEMBER).
✅ 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:v0.5.17-rocm720-mi35x-20260818 is the upstream SGLang org on established MI355X; the entry is itself framework: sglang, so ordering does not apply.
✅ Check 7 (deprecations): PASS — qwen3.5 fp4 agentic-coding (MTP arm) is active per MODELS.md on 2026-08-20.
✅ Check 8 (no architecture hacks): PASS — the only server-arg change is --kv-cache-dtype fp8_e4m3 (lower-precision KV, evals pass); no --hf-overrides or FLOPs reduction.
✅ Check 9 (chat template): PASS — the agentic replay adds --apply-chat-template in qwen3.5_fp4_mi355x_sglang_mtp.sh.
✅ Check 10 (no engine patches): PASS — the diff introduces no patching of the pinned image.
✅ Check 11 (golden AL): PASS — SGLANG_SIMULATE_ACC_LEN=3.39 with match-expected / real-draft-token matches the committed golden AL (golden_al_distribution/qwen3.5_mtp.yaml, thinking_on, 3 speculative tokens → 3.39); eval keeps real verification.
➖ Check 12 (append-only): N/A — the new perf-changelog.yaml entry does not use append-only: true.