[shyeh25]

Summary

Update the Muse Spark Open B300 SGLang AgentX curve to lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc and retune the per-concurrency points:

• TP8/EP8, GPU-resident KV: concurrency 1, 4, 7, 16, 20, 24
• TP4/EP4 on 4 GPUs, DRAM HiCache: concurrency 16, 28, 32, 38, 48

TP8 concurrency 8, 10, 14 and TP4 HiCache concurrency 40, 44 are removed.

Every TP4 HiCache point disables prefill CUDA graphs and lowers mem-fraction-static from 0.95 to 0.92. cuda-graph-backend-prefill: breakable moves from the base to each TP8 point; TP8 concurrency 1 keeps its settings.

Per-point changes:

• TP8 concurrency 4, starting from the concurrency 1 settings:
  • request admission and decode graph batch 2 → 1, no prefill-decode interval
  • shortest-prefill-first scheduling
  • stream interval 20 → 40
• TP8 concurrency 7 (new), starting from the previous concurrency 8 settings:
  • shortest-prefill-first scheduling
  • stream interval 40
• TP8 concurrency 16:
  • request admission and decode graph batch 4 → 3
  • prefill-decode interval 64 → 88
  • stream interval 40 with incremental streaming output
• TP8 concurrency 20: prefill-decode interval 64 → 48
• TP8 concurrency 24: prefill-decode interval 64 → 80
• TP4 concurrency 16: scheduler receive interval 20 and stream interval 20
• TP4 concurrency 28: prefill-decode interval 128 → 104
• TP4 concurrency 32:
  • prefill-decode interval 128 → 96
  • HiCache 170 → 250 GB per rank, SWA/full token ratio 0.1 → 0.2
  • SGLANG_TRTLLM_MHA_DECODE_SEQ_LEN_SPLITS=4
• TP4 concurrency 38 (new), starting from the previous concurrency 40 settings:
  • request admission 80 → 16, decode graph batch 80 → 76
  • prefill-decode interval 96 → 56
  • HiCache 170 → 250 GB per rank
  • SGLANG_DFLASH_EAGER_DRAFT_SAMPLER=0
• TP4 concurrency 48: only the TP4-wide prefill CUDA graph and memory fraction change

Chunked prefill, the checkpoint, and the DFlash2 draft settings are unchanged.

AI model disclosure

• Model/version: claude-opus-5-5
• Role: tuning, config changes

PR checklist and change type

Type of Change

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other (please describe)

Checklist

• I have completed the AI model disclosure and kept it current
• I have tested my changes locally
• I have updated documentation if necessary
• If python -m infx.matrix.changed --base origin/main lists affected config keys, I have appended one new entry covering them to the physical end of inferencex-e2e/perf-changelog.yaml; if it lists none, I have added none. I have not edited historical entries
• If this PR appends an inferencex-e2e/perf-changelog.yaml entry, it carries exactly one primary sweep label (a maintainer applies it on fork PRs): full-sweep-fail-fast (recommended), full-sweep-enabled, or non-canary-full-sweep-enabled. Optional modifiers all-evals, evals-only, and agentx-fast require a primary label; the last two block reuse while applied.
• Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /use <run_id> (or the legacy /reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.

中文

将 Muse Spark Open B300 SGLang AgentX 曲线更新到 lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc，并按并发重新调优各点：

• TP8/EP8，KV 留在 GPU：并发 1、4、7、16、20、24
• TP4/EP4（4 张 GPU），DRAM HiCache：并发 16、28、32、38、48

移除 TP8 并发 8、10、14 与 TP4 HiCache 并发 40、44。

所有 TP4 HiCache 点都关闭 prefill CUDA graph，并将 mem-fraction-static 从 0.95 降到 0.92。cuda-graph-backend-prefill: breakable 从 base 移到各 TP8 点；TP8 并发 1 设置不变。

各点改动：

• TP8 并发 4，以并发 1 的设置为基础：
  • 请求数上限与 decode graph batch 2 → 1，不设 prefill-decode interval
  • shortest-prefill-first 调度
  • stream interval 20 → 40
• TP8 并发 7（新增），以原并发 8 的设置为基础：
  • shortest-prefill-first 调度
  • stream interval 40
• TP8 并发 16：
  • 请求数上限与 decode graph batch 4 → 3
  • prefill-decode interval 64 → 88
  • stream interval 40，并启用 incremental streaming output
• TP8 并发 20：prefill-decode interval 64 → 48
• TP8 并发 24：prefill-decode interval 64 → 80
• TP4 并发 16：scheduler receive interval 20，stream interval 20
• TP4 并发 28：prefill-decode interval 128 → 104
• TP4 并发 32：
  • prefill-decode interval 128 → 96
  • HiCache 每 rank 170 → 250 GB，SWA/full token 比例 0.1 → 0.2
  • SGLANG_TRTLLM_MHA_DECODE_SEQ_LEN_SPLITS=4
• TP4 并发 38（新增），以原并发 40 的设置为基础：
  • 请求数上限 80 → 16，decode graph batch 80 → 76
  • prefill-decode interval 96 → 56
  • HiCache 每 rank 170 → 250 GB
  • SGLANG_DFLASH_EAGER_DRAFT_SAMPLER=0
• TP4 并发 48：只有 TP4 共通的 prefill CUDA graph 与 memory fraction 改动

chunked prefill、checkpoint 与 DFlash2 draft 设置不变。

AI 模型使用说明：claude-opus-5-5；工作内容：调优、配置改动。

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Claude finished @shyeh25's task in 25s —— View job

---

Re-review: LGTM, no blocking issues found. Commit 7516df00a fixed the earlier pr-link: .../pull/TBD finding, and the entry now points to #4725. The recipe overrides match the master-config conc-list values (TP8: 1, 4, 7, 16, 20, 24; TP4 HiCache: 16, 28, 32, 38, 48). The image is public on Docker Hub, and the changelog entry is appended at the end of the file. I could not run infx.matrix.changed here because the command needed approval, but the PR already appends an entry for musespark-fp4-b300-sglang-agentic-dflash, the only config key it touches.

[claude]

I reviewed the latest push and found no new bugs. The changelog pr-link now points to a real PR (#4725), but a human look at the retuned per-concurrency settings is still worthwhile.

What was reviewed:

• Every conc in the master conc-lists (TP8: 1, 4, 7, 16, 20, 24; TP4: 16, 28, 32, 38, 48) has a matching override_tpN_cM key in agentic.yaml, and no stale c8/c10/c14/c40/c44 keys remain.
• cuda-graph-backend-prefill: breakable moved from the base into every TP8 override. Every TP4 override now has mem-fraction-static 0.92 and disable-prefill-cuda-graph.
• The new perf-changelog block is appended only at the tail, and no existing bytes changed.

Extended reasoning...

The change touches one single-node srt-slurm recipe, the matching nvidia-master.yaml entry (image bump plus conc-list changes) and one appended perf-changelog entry. None of it touches a security-sensitive surface. The tuning changes affect only benchmark config, but they include draft-related settings (speculative-draft-attention-backend fa4 at TP8 c4, SGLANG_DFLASH_EAGER_DRAFT_SAMPLER=0 at TP4 c38). The repo's speculative-decoding rules call for closer scrutiny of settings like these. TP4 c38 also sets max-running-requests 16 with cuda-graph-max-bs-decode 76, unlike the matched pairs in the other overrides.

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=41328422543

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=41328422543

[nvpohanh]

/use 41328422543

[github-actions]

@nvpohanh staged run 41328422543.

• Muse-Spark-Open · agentic-traces · FP4 · b300_sglang · 2026-11-01

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

[nvpohanh]

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (musespark-fp4-b300-sglang-agentic-dflash) will happen in sgl-project/sglang#47929, in the Agentic Long-Context Deployment section of the Muse Spark Open cookbook page.

The commands mirror the launch measured in run 41328422543, pinned by this PR's /use comment.

[nvpohanh]

[by Claude Code] sgl-project/sglang#47929 has been merged. This PR's AgentX config (musespark-fp4-b300-sglang-agentic-dflash) is now in the Agentic Long-Context Deployment section of the SGLang Muse Spark Open cookbook: https://docs.sglang.io/cookbook/autoregressive/Meta/MuseSparkOpen#agentic-long-context-deployment

The commands mirror the launch measured in run 41328422543, pinned by this PR's /use comment.

[nvpohanh]

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

Checklist items

• 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.
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• Verified that every draft model and draft head is served as it ships: the draft that ships with the served checkpoint, at its stored precision, through the pinned upstream image's default handling, with the shipped and effective draft precision recorded in the additional detail section. No submission-side quantization, dtype override, checkpoint substitution, or patch may lower draft precision below that default, regardless of eval results or AL. Explicitly verified that SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not enabled in the effective recipe, including inherited settings; enabling it is prohibited going forward, and historical runs do not grant an exception. See Draft-model precision for what counts as the default and the MLPerf comparison.
• 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 infx/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; target/verifier FLOPs at lower precisions is fine, given that the config passes private evals, but this does not permit lowering draft-model or draft-head precision below what ships. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
• Verified that new models disable integer all-reduce, with SGLang ROCm explicitly setting ROCM_QUICK_REDUCE_QUANTIZATION=NONE. Existing models are grandfathered per MODELS.md.
• For affected Dynamo TRT-LLM benchmarks, verified that iteration/KV-cache metrics are collected and published under the pinned Dynamo implementation and effective engine settings. Resolve the installed Dynamo version separately from the container tag; use supported metrics controls without requiring KV events. Record the version and configuration evidence in the additional detail section. See verifier Check 17. N/A for unaffected serving paths or frameworks without these Dynamo settings.
• 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
• For official InferenceX (non-preview), verified that established models use upstream engine images built from master (or the equivalent default development branch), with new models transitioning within a reasonable time after release. See MODELS.md.
• 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 affected vLLM/SGLang production serving configuration in this PR, including Dynamo deployments, is documented in the official vLLM recipes and/or the SGLang cookbook, for both single-node and multi-node aggregated and disaggregated serving. The documentation covers production-useful flags, environment variables, and launch commands (including prefill, decode, and router/frontend commands for disaggregation); InferenceX-only benchmark/harness settings are excluded with reasoning in the additional detail section:
  • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang cookbook source 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.
• Reported measured throughput/E2EL Pareto counts and evidence per affected curve (≥5 points strongly recommended). Below 5 or unverifiable: tag a core maintainer for review; recorded admin bypass required before merge. N/A if no curves are affected. Details.

Additional detail section:

• Scope: one config, musespark-fp4-b300-sglang-agentic-dflash (single-node aggregated SGLang on B300 with DFlash2), agentic-coding only. Changed files: musespark/sglang/b300-fp4-mtp/agentic.yaml, nvidia-master.yaml (image nightly-dev-20261021-c1c287a4 → nightly-dev-cu13-20261030-302b78cc; TP8/EP8 c1, 4, 7, 16, 20, 24 and TP4/EP4 HiCache c16, 28, 32, 38, 48) and one appended perf-changelog.yaml entry. Assessed head: 7516df00. The PR has two commits, both non-merge, and the pinned run is on the head.
• PR validation: Run Sweep 41328422543 (attempt 1, pull_request) on head 7516df00. All 11 agentic / jobs succeeded (TP8 c1, 4, 7, 16, 20, 24; TP4 HiCache c16, 28, 32, 38, 48), as did all 5 agentic eval / jobs, check-changelog and collect-evals. Fixed-sequence and multi-node lanes were skipped because this PR has no such configs. The run is pinned with /use 41328422543, and its bmk_agentic_* and eval_results_all artifacts are unexpired (until 2027-01-30). Every point used the semianalysis_cc_traces_weka_062126 corpus.
• Evals: eval_results_all from run 41328422543, gsm8k em_strict (n=1319): TP4 c16 0.9788, c28 0.9757, c38 0.9780, c48 0.9765; TP8 c24 0.9765. All are above the 0.90 default threshold in infx/evals/thresholds.yaml (no musespark override). They are within 2σ of each other and of the closest ingested B300 cohort (run 40285747600, TP4 c48, 0.9757). The evals ran on the same image as the benchmark jobs.
• Chat template: AgentX replays through agentx profile with AIPerf input_config.endpoint.type: chat (streaming chat completions) at every point (profile_export_aiperf.json in each server_logs_* artifact). Every server runs with chat_template='/model/chat_template.jinja'.
• Draft precision (served as shipped): the draft is RadixArk/Muse-Spark-Open-DFlash2 (speculative_draft_model_revision='main', HF sha 46c90f41). It ships as BF16: safetensors metadata reports 4,887,479,552 parameters, all BF16. No point sets a draft quantization, dtype, revision, checkpoint or draft-KV-dtype override beyond what is noted below. On every point the server args show _speculative_draft_quantization_explicitly_set=False and json_model_override_args='{}'.
  • The server args show speculative_draft_model_quantization='modelopt_fp4'. This comes from the pinned image's default inheritance of the target's --quantization (serving_hook.py), not from a submission setting. The draft checkpoint has no quantization metadata, so weight_utils.get_quant_config uses ModelOptFp4Config.for_online_weight_quantization → nvfp4_online.get_quant_method. That returns UnquantizedLinearMethod for every LinearBase layer (the checkpoint is not FP8-serialized) and quantizes only FusedMoE. models/dflash.py (DFlash2DraftModel, dense DFlashMLP) has no FusedMoE. This holds at the pinned image's sglang commit.
  • The logs match: on every point the draft loads as Load weight end … type=DFlash2DraftModel with no quant tag, at 2.60 GB per GPU at TP8 and 3.63 GB per GPU at TP4 (BF16-sized). The target logs quant=modelopt_fp4, quant_algo=NVFP4. Shipped draft precision: BF16. Effective draft weight/activation precision: BF16, at every point.
  • Draft KV cache: TP8 c1 and c4 use fa4 draft attention, and the image forces BF16 draft KV there (DFLASH fa4 draft: overriding KV cache dtype torch.float8_e4m3fn -> torch.bfloat16). Every other point leaves the draft attention backend unset, so the draft follows the image default (trtllm_mha, the target backend) and inherits the target's upstream-supported fp8_e4m3 KV dtype (draft pool KV Cache is allocated. dtype: torch.float8_e4m3fn or Post-capture KV sizing: draft KV cache allocated. dtype: torch.float8_e4m3fn). No point sets speculative-draft-kv-cache-dtype. SGLANG_DFLASH_EAGER_DRAFT_SAMPLER (base 1, 0 at TP4 c38) only decides whether the draft greedy head runs eagerly or inside the draft CUDA graph (dflash_worker_v2.py::_maybe_build_draft_sampler). It does not change draft precision.
  • SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 is already in this recipe's base env on main (unchanged by this PR). It is a target-only setting. At the pinned sglang commit it is read only by ModelOptFp4Config.__init__ for serialized NVFP4 checkpoints and by the MoE runner hook. It selects per-tensor instead of per-token FP32 activation scales for the target's NVFP4 GEMMs, and the passing evals cover it. The draft has no serialized NVFP4 metadata (online config, see above), so it is unaffected.
  • SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not enabled. It does not appear in the recipe, the master config, the generated config.yaml/sbatch_script.sh, or the server logs of any point. Its pinned default is EnvBool(False) (environ.py), and DFlash2 is a dense draft with no NextN MoE.
• Golden AL: DFlash speculative-num-draft-tokens: 8 maps to num_speculative_tokens = 7. musespark_dflash.yaml gives muse-spark-open-816b-a42b-nvfp4 / thinking_on / 7 = 4.05 (signoff_facts golden-al --model musespark --method dflash --tokens 7 --thinking thinking_on → matches: true). Every point's server env has SGLANG_SIMULATE_ACC_LEN=4.05, SGLANG_SIMULATE_ACC_METHOD=match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token, injected by the harness (the recipe hard-codes none).
• MODELS.md: musespark (Muse Spark Open 816B-A42B, added 2026-10-15) lists Agentic coding with SGLang and the DFlash2 draft as active, with no deprecated scenarios (review date 2026-11-03).
• No architecture hacks: there is no --hf-overrides or json-model-override-args (effective '{}'), and no layer, expert or indexer skipping. The changes cover image, request admission, CUDA-graph batch/backend, memory fraction, SWA pool ratio, scheduling policy, prefill-decode interval, stream interval, chunked prefill, HiCache sizing, and draft attention backend.
• Integer all-reduce: musespark is in the policy-baseline model support matrix, so it is grandfathered. This is also a CUDA path: no *QUICK_REDUCE_QUANTIZATION or other all-reduce quantization control appears in the recipe, the resolved configs, or the captured worker environment.
• Dynamo TRT-LLM metrics: N/A. This is a direct SGLang deployment (framework: sglang) with no Dynamo TRT-LLM serving path.
• Upstream image / main-branch provenance / engine order: lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc comes from the upstream lmsysorg org and resolves now (check_18c_images). On B300 (amd64), the run fingerprint reports SGLang 0.0.0.dev1+g302b78ccf. sgl-project/sglang 302b78cc (#46024) is an ancestor of main (compare 302b78cc...main: ahead 325, behind 0), so this is a main-branch nightly. This is an SGLang submission, so the engine-first ordering is satisfied.
• Upstream recipe: sgl-project/sglang#47929 is MERGED (2026-11-03T03:28:14Z, merge commit cd41d53d69). It is published at https://docs.sglang.io/cookbook/autoregressive/Meta/MuseSparkOpen#agentic-long-context-deployment (docs/src/snippets/agentx/meta-models/muse-spark-open.jsx), with B300 cells musespark-fp4-b300-sglang-agentic-dflash#0 (Agg TP8·EP8) and #1 (Agg TP4·EP4, HiCache), provenance sha 7516df00, run 41328422543, pr 4725.
  • I decoded every cookbook launch with the page's own decoder (_agentx.jsx at sglang main) and compared it, point by point, with the resolved config.yaml and the actual python3 -m sglang.launch_server argv in each server_logs_* artifact of run 41328422543. Every flag and env var matches for every measured point: image, model, TP/EP, modelopt_fp4, flashinfer_trtllm MoE, fp8_e4m3 KV, DFLASH + DFlash2 with 8 draft tokens, draft attention backend, muse parsers, and all per-concurrency overrides (admission, CUDA-graph batch/backend, memory fraction, SWA ratio, scheduling, prefill-decode interval, stream interval, chunked prefill, HiCache size/write policy, layernorm SP/MegaMoE, and the SGLANG_*/PYTORCH_CUDA_ALLOC_CONF env). The cookbook additionally shows NCCL_CUMEM_ENABLE=1, which the measured workers also had (cluster environment).
  • InferenceX-only settings excluded from upstream coverage: --enable-metrics (scraped by the AIPerf harness); --enable-cache-report (adds cached-token counts to response usage for the harness's cache accounting, with no change to scheduling, caching or kernels); the staged /model path, --host/--port/--nccl-port launcher plumbing; PYTHONNOUSERSITE/PYTHONUNBUFFERED; HF_HUB_OFFLINE; health-check polling; all benchmark.env client settings (TOKENIZER, AIPERF_*, CONC, KV_OFFLOADING); and the harness-injected synthetic-acceptance env (SGLANG_SIMULATE_ACC_*).
  • NVIDIA overview: Muse-Spark-Open is listed with SGLang ✅ in the Framework Recipes table of https://docs.nvidia.com/llm-inference-quick-start-recipes/overview.
• No engine patches: the diff changes only YAML. The effective recipe has no setup script, patch, pip install, file overwrite or monkey-patching, and the generated sbatch_script.sh and config.yaml of every point contain none. No waiver is needed.
• append-only: true: N/A. The appended changelog entry does not use append-only.
• Pareto coverage: there is one affected curve: Muse-Spark-Open · agentic-coding · b300_sglang · FP4. All points of this config share that series, and no other config maps to it. Counts use the bmk_agentic_* artifacts of run 41328422543 (attempt 1, head 7516df00), with x = E2EL (s) and y = request_metrics.throughput.per_gpu.total_tput_tps, computed with infx.workflows.pareto_coverage:
  • P90 E2EL: 8 frontier points / 11 measured (TP8 c7, c20, c24; TP4 HiCache c16, c28, c32, c38, c48).
  • P75 E2EL (reported separately): 8 frontier points / 11 measured (TP8 c7, c20, c24; TP4 HiCache c16, c28, c32, c38, c48).
  • No invalid or missing points.
• Removed points: TP8 c8, c10, c14 and TP4 c40, c44. signoff_facts report at 7516df00 finds undominated_removed: [] and unmeasured_removed: [], so every removed point is covered by the measured head curve. Re-tuned deltas (tput/GPU, mean interactivity): TP4 c32 +21.6% / −29.3%, TP8 c16 −1.3% / +28.8%, TP8 c4 −7.6% / +19.6%, TP4 c16 −1.2% / +11.1%; the rest are within ±5%.
• CPU DRAM policy: the largest HiCache pool is 250 GB per rank × 4 ranks = 1.0 TB at TP4 (4 of 8 GPUs), below the 1.5 TB proportional cap for HGX B300 in MODELS.md.
• The PR branch currently conflicts with main only in the append-only inferencex-e2e/perf-changelog.yaml tail (git merge-tree), which is resolved at merge. The pinned run stays reusable for that merge.

Signed: nvpohanh

[github-actions]

✅✅✅ Verdict: PASS ✅✅✅

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — @nvpohanh is a named owner of inferencex-e2e/configs/nvidia-master.yaml and of /inferencex-e2e/benchmarks/*_node/srt-slurm-recipes/**/b300-*/; perf-changelog.yaml falls under the catch-all, which any recognized CODEOWNER covers.

✅ Check 1 (Sweep on in-PR commit): PASS — Run Sweep 41328422543 (attempt 1, pull_request) ran on the pinned head 7516df00, which is still the PR tip. All 11 executed agentic / jobs and all 5 agentic eval / jobs succeeded; the fixed-sequence single-node */ and eval / lanes were skipped because this agentic-only config has none. No commits follow the validated SHA.

✅ Check 2 (Evals): PASS — eval_results_all from run 41328422543 reports gsm8k em_strict of TP4 c16 0.9788, c28 0.9757, c38 0.9780, c48 0.9765 and TP8 c24 0.9765, all with n=1319 on lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc. Every row is within 2σ of its best sibling and of the exact B300 cohort (40285747600 TP4 0.9757; 39181803276 TP8 0.9773). The change is a per-point retune plus an image bump, with no conditional kernel path, and the evals ran on the bumped image across both topologies.

✅ Check 3 (Upstream recipe): PASS — sgl-project/sglang#47929 is MERGED (2026-11-03T03:28:14Z, cd41d53d). At sglang main, docs/src/snippets/agentx/meta-models/muse-spark-open.jsx carries musespark-fp4-b300-sglang-agentic-dflash#0 (TP8·EP8) and #1 (TP4·EP4 HiCache) with provenance sha 7516df00 / run 41328422543. Its base launch and per-point patches match the pinned recipe for every point: image, model, TP/EP, modelopt_fp4, flashinfer_trtllm, fp8_e4m3 KV, DFLASH with 8 draft tokens, fa4 draft attention at c1/c4, admission and graph batch, prefill-decode interval, stream interval, SWA ratio, chunked prefill, HiCache size and write policy, layernorm SP/MegaMoE, prefill CUDA-graph settings, and the SGLANG_* env. The sign-off justifies its exclusions as harness-only: metrics, cache report, launcher plumbing, client env, and synthetic acceptance.

✅ Check 4 (Reuse command): PASS — nvpohanh (COLLABORATOR) posted /use 41328422543 on its own line; that is the passing in-PR run from Check 1.

✅ Check 5 (Latest template): PASS — The sign-off contains every item of the current PR_REVIEW_CHECKLIST.md template, including the sub-item, all checked, with the additional detail section filled in.

✅ Check 6 (Upstream image / provenance / engine order): PASS — lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc comes from the upstream lmsysorg org. sglang 302b78ccf is an ancestor of main (compare: ahead 326, behind 0), so it is a main-branch nightly. The PR adds only a framework: sglang entry, so engine-first ordering is satisfied.

✅ Check 7 (Deprecated models): PASS — musespark (added 2026-10-15) lists Agentic coding as active in MODELS.md with no deprecated scenarios as of 2026-11-03.

✅ Check 8 (No architecture hacks): PASS — The PR has no --hf-overrides, json-model-override-args, or layer/expert/indexer skipping, and no accuracy-trading lever needs extra eval coverage. musespark appears in the policy-baseline support matrix, so it is grandfathered. This is also a CUDA path, and no *QUICK_REDUCE_QUANTIZATION control appears in the recipe or the shared launchers.

✅ Check 9 (Chat template): PASS — The agentic replay runs agentx profile over the semianalysis_cc_traces_weka_062126 chat-message traces, using the shared, unchanged harness path (infx/bench/agentic/replay.py). The PR does not change the client.

✅ Check 10 (No engine patches): PASS — The diff changes only YAML. The pinned recipe has no setup_script, patch, pip install, or file overwrite. Its srtctl overrides in srt-single-node-submission.json add only harness env, eval wiring, and SGLANG_SIMULATE_ACC_*.

✅ Check 11 (Golden AL): PASS — speculative-num-draft-tokens: 8 means 7 speculative tokens, and musespark_dflash.yaml thinking_on / 7 gives 4.05. The run's applied overrides set SGLANG_SIMULATE_ACC_LEN=4.05, SGLANG_SIMULATE_ACC_METHOD=match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token on the agg role.

➖ Check 12 (Append-only): N/A — The new perf-changelog.yaml entry does not set append-only: true.

✅ Check 13 (Draft precision): PASS — The draft RadixArk/Muse-Spark-Open-DFlash2 ships as dense BF16: config.json has dtype: bfloat16, no quantization_config, and no MoE fields. At sglang 302b78ccf, a modelopt_fp4 draft without quant metadata takes ModelOptFp4Config.for_online_weight_quantization (weight_utils.py). There, get_quant_method returns UnquantizedLinearMethod for every non-FP8-serialized LinearBase and quantizes only FusedMoE, and models/dflash.py contains no FusedMoE, so the effective draft weights stay BF16. Draft KV is fp8 inherited from the target, or the image forces BF16 for fa4. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is set nowhere in the recipe or shared launchers.

✅ Check 14 (Pareto coverage): PASS — Recomputed from the 11 bmk_agentic_* artifacts of run 41328422543 (attempt 1, 7516df00) for Muse-Spark-Open / agentic-coding / b300_sglang / FP4, using x = E2EL and y = per_gpu.total_tput_tps. P90 E2EL has 8 frontier points (TP8 c7, c20, c24; TP4 c16, c28, c32, c38, c48), and P75 E2EL, reported separately, also has 8.

✅ Check 15 (PR description): PASS — The PR body's per-point claims match the pinned recipe and the master conc-list values: c4/c7/c16/c20/c24 at TP8; c16/c28/c32/c38/c48 at TP4; the image; and the removed TP8 c8/c10/c14 and TP4 c40/c44 points. Its validation claims match run 41328422543.

✅ Check 16 (NVIDIA overview): PASS — The NVIDIA overview Framework Recipes table lists Muse-Spark-Open with SGLang ✅, covering this B300 SGLang pair.

➖ Check 17 (Metrics publication): N/A — This is a direct SGLang deployment (framework: sglang) with no Dynamo TRT-LLM path.

✅ Check 18 (Recipe/image/sweep consistency): PASS — Both config keys generate at head, and the master TP/EP/GPU counts and conc-list match the recipe overrides. The new image resolves (check_18c_images), and the PR changes no orchestration or dispatch. The changelog accurately describes the image, points, prefill-graph/memory, admission, interval, HiCache and SWA changes. undominated_removed and unmeasured_removed are both empty, and spot checks confirm it: TP4 c44 lies under the c38–c48 segment, and TP8 c8/c10/c14 are dominated by c16. Re-tune deltas: TP4 c32 +21.6% tput / −29.3% intvty, TP8 c16 −1.3% / +28.8%, TP8 c4 −7.6% / +19.6%, TP4 c16 −1.2% / +11.1%; all others are within ±5%.

Assessed commit: 7516df00ac37c119055834dbc6d48811ae6a0b99.

[functionstackx]

Approved at Oren’s request.

[nvpohanh]

/reuse-sweep-run 41328422543

[github-actions]

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category	new	duplicate	skipped	unmapped	failed
benchmark	11	0	0	0	0
eval	5	5	5	0	0
trace	11	0	0	0	0
power	11	0	0	0	0

_{benchmarkFiles 11, rawBenchmarkRows 11, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 6595, gpuMetricSamplesNew 386092, powerPublicationPoints 11, changelogEntriesWritten 1}

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据，各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。