Skip to content

[AgentX] Retune Muse Spark Open GB200 SGLang per concurrency / [AgentX] 按并发重新调优 Muse Spark Open GB200 SGLang - #4670

Merged
nvpohanh merged 4 commits into
mainfrom
musespark-fp4-gb200-sglang-agentic-retune-1029
Nov 6, 2026
Merged

nvpohanh merged 4 commits into
mainfrom
musespark-fp4-gb200-sglang-agentic-retune-1029

Conversation

@shyeh25

@shyeh25 shyeh25 commented Oct 30, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Retune the Muse Spark Open GB200 SGLang AgentX curve on lmsysorg/sglang:nightly-dev-20261026-8252e3af (image unchanged) and set SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 in the base serving environment:

  • TP4/EP4, GPU-resident KV: concurrency 1 (unchanged)
  • TP4/EP4, DRAM HiCache: concurrency 4, 7, 10, 14, 16, 28, 36, 40

HiCache concurrency 2, 6, 8, 20, 24 and 32 are removed. Every removed point that was on the previous GB200 frontier is beaten on both throughput per GPU and interactivity by a retained point (6 and 8 by concurrency 10, 20 and 24 by concurrency 28, 32 by concurrency 40); concurrency 2 was already dominated by concurrency 4.

Per-point changes:

  • Concurrency 4:
    • prefill-decode interval 128 → 256
    • shortest-prefill-first scheduling
    • stream interval 20
  • Concurrency 7 (new), starting from the previous concurrency 4 settings:
    • prefill-decode interval 64
    • stream interval 20
  • Concurrency 10:
    • request admission and decode graph batch 10 → 3
    • prefill-decode interval 32 → 64
    • stream interval 20, 4 tokenizer workers, incremental streaming output
  • Concurrency 14 (new), starting from the concurrency 10 settings:
    • request admission and decode graph batch 5
    • prefill-decode interval 96
  • Concurrency 16:
    • request admission and decode graph batch 32 → 6
    • prefill-decode interval 16 → 64
    • stream interval 20, 4 tokenizer workers, incremental streaming output
  • Concurrency 28:
    • memory fraction 0.9
    • chunked prefill and max prefill tokens 16384
    • shortest-prefill-first scheduling
    • prefill-decode interval 16 → 96
    • stream interval 20
    • SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=4096, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
  • Concurrency 36 (new), starting from the concurrency 28 settings:
    • request admission and decode graph batch 72
    • prefill-decode interval 128
  • Concurrency 40 (new), starting from the concurrency 28 settings:
    • request admission and decode graph batch 48
    • memory fraction 0.93
    • HiCache size 160 GB
    • prefill-decode interval 64

The checkpoint, DFlash2 draft settings, the concurrency 1 server settings and the image are unchanged.

AI model disclosure
  • Model/version: claude-opus-5-5.
  • Role: claude-opus-5-5 directed the tuning campaign, analysed the measured runs, generated the recipe, master-config and perf-changelog changes, checked them locally, and drafted this PR description; the ARIA orchestrating agent and its task agents (gpt-6.1-sol via Codex) submitted and collected the benchmark and GSM8K runs.
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

在 lmsysorg/sglang:nightly-dev-20261026-8252e3af(镜像不变)上重新调优 Muse Spark Open GB200 SGLang AgentX 曲线,并在基础服务环境中设置 SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0:

  • TP4/EP4,KV 仅驻留 GPU:并发 1(不变)
  • TP4/EP4,DRAM HiCache:并发 4、7、10、14、16、28、36、40

移除 HiCache 并发 2、6、8、20、24、32。原 GB200 前沿上被移除的点,都被保留的点在每 GPU 吞吐和交互性两个维度同时超过(6 和 8 被并发 10 超过,20 和 24 被并发 28 超过,32 被并发 40 超过);并发 2 原本已被并发 4 支配。

各点改动:并发 4 将 prefill-decode interval 由 128 提高到 256,并使用 shortest-prefill-first 与 stream interval 20;新增并发 7(请求准入 2、prefill-decode interval 64、stream interval 20);并发 10 与 16 将请求准入与 decode graph 批大小分别降到 3 和 6,prefill-decode interval 提高到 64,并使用 stream interval 20、4 个 tokenizer worker 和增量流式输出;新增并发 14(准入 5、interval 96);并发 28 使用显存占用比例 0.9、16384 token chunked prefill、shortest-prefill-first、interval 96、stream interval 20、SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=4096;新增并发 36(准入 72、interval 128);新增并发 40(准入 48、显存占用比例 0.93、HiCache 160 GB、interval 64)。模型检查点、DFlash2 草稿模型设置、并发 1 的服务端设置和镜像保持不变。

AI 模型使用说明:模型为 claude-opus-5-5。claude-opus-5-5 负责指挥调优、分析实测结果、生成配方、主配置和性能变更记录的改动、在本地检查,以及起草本 PR 说明;ARIA 编排代理及其任务代理(gpt-6.1-sol,经 Codex)负责提交与收集性能和 GSM8K 测试。

🤖 Generated with Claude Code

Retune concurrency 4, 10, 16 and 28, add concurrency 7, 14, 36 and 40,
drop concurrency 2, 6, 8, 20, 24 and 32, and set
SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 in the base environment.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

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 Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I reviewed this PR and didn't find any bugs. The recipe and master config are both under NVIDIA code ownership, and the per-concurrency tuning values are benchmark choices a maintainer should sign off on, so a human look is still worthwhile.

What was reviewed:

  • The conc-list in nvidia-master.yaml (4, 7, 10, 14, 16, 28, 36, 40) matches the override_tp4_cN keys in agentic.yaml. Each HiCache override now sets KV_OFFLOADING: dram and has a matching CONC.
  • The new changelog entry matches the recipe: SPF everywhere except c7, admission of 2-6 at c4-16, and 16K chunked prefill at c28-40. It is appended at the tail with no historical bytes changed. The pr-link is still the usual XXX placeholder.
  • The c1 settings, the image and the DFlash2 draft settings are unchanged.
Extended reasoning...

The change retunes one benchmark config (musespark-fp4-gb200-sglang-agentic-dflash). It touches its single-node srt-slurm recipe, a single conc-list line in nvidia-master.yaml and an appended perf-changelog entry. It touches nothing security-sensitive. All three files fall under CODEOWNERS rules for NVIDIA recipes, the NVIDIA master config or the core team, so this shouldn't be auto-approved. The new per-point serving values are empirical tuning choices that only the benchmark results can confirm.

@shyeh25 shyeh25 added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 30, 2026
@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (musespark-fp4-gb200-sglang-agentic-dflash) will happen in sgl-project/sglang#47464, in the Agentic Long-Context Deployment section of the Muse Spark Open cookbook page.

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (5692839d9) as "in progress". A later cookbook PR switches it to the measured run once a run is pinned or this PR merges.

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] sgl-project/sglang#47464 has been merged. This PR's AgentX config (musespark-fp4-gb200-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

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (5692839d9) as "in progress". A later cookbook PR switches it to the measured run once a run is pinned or this PR merges.

@github-actions

github-actions Bot commented Oct 30, 2026 •

Copy link
Copy Markdown
Contributor

shyeh25 and others added 2 commits October 30, 2026 10:00
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nvpohanh

nvpohanh commented Nov 2, 2026

Copy link
Copy Markdown
Collaborator

/use 41096285400

@github-actions

github-actions Bot commented Nov 2, 2026

Copy link
Copy Markdown
Contributor

@nvpohanh staging run 41096285400. Existing staged runs will be preserved; a completion link will be posted here.

@nvpohanh

nvpohanh commented Nov 2, 2026

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (musespark-fp4-gb200-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 config has been live in that section since sgl-project/sglang#47464; sgl-project/sglang#47929 refreshes it with the pin below.

The commands mirror the launch measured in run 41096285400, pinned by this PR's /use comment.

@nvpohanh

nvpohanh commented Nov 3, 2026

Copy link
Copy Markdown
Collaborator

[by Claude Code] sgl-project/sglang#47929 has been merged. This PR's AgentX config (musespark-fp4-gb200-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 41096285400, pinned by this PR's /use comment.

@nvpohanh

nvpohanh commented Nov 3, 2026

Copy link
Copy Markdown
Collaborator

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-gb200-sglang-agentic-dflash (single-node aggregated SGLang on GB200 with DFlash2), agentic-coding only. Changed files: musespark/sglang/gb200-fp4-mtp/agentic.yaml (base env SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 plus per-concurrency overrides), nvidia-master.yaml (HiCache conc-list [2, 4, 6, 8, 10, 16, 20, 24, 28, 32] → [4, 7, 10, 14, 16, 28, 36, 40]; image unchanged) and one appended perf-changelog.yaml entry. Assessed head: c9eb455f, a merge of main into d75601d5. The pinned run is on this head.
  • PR validation: Run Sweep 41096285400 (attempt 1, pull_request) on head c9eb455f. The canary (TP4 c1) and all 9 agentic / jobs succeeded (TP4 c1; TP4 HiCache c4, 7, 10, 14, 16, 28, 36, 40), as did all 4 agentic eval / jobs, check-changelog, collect-evals, collect-results and calc-success-rate. Fixed-sequence and multi-node lanes were skipped because this PR has no such configs. The run is pinned with /use 41096285400, and its results_bmk, bmk_agentic_* and eval_results_all artifacts are unexpired (until 2027-01-28). Every point used the semianalysis_cc_traces_weka_062126 corpus.
  • Evals: eval_results_all from run 41096285400, gsm8k em_strict (n=1319): TP4 c1 0.9795, HiCache c7 0.9795, c16 0.9765, c40 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 GB200 cohort (run 40279195159, TP4 c32, 0.9757). The evals ran on the same image as the benchmark jobs, with the new SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 base env in effect at every eval point.
  • 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 3.61–3.63 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: c1, c4 and c7 use fa4 draft attention, and the image forces BF16 draft KV there (overriding KV cache dtype torch.float8_e4m3fn -> torch.bfloat16). c10–c40 unset the draft attention backend, 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 allocated as torch.float8_e4m3fn). No point sets speculative-draft-kv-cache-dtype; the earlier explicit FP8 draft KV at c40 was dropped in d75601d5.
    • SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 (new in this PR's base env) is a target-only setting. At sglang 8252e3af it is read only by ModelOptFp4Config.__init__ for serialized NVFP4 checkpoints (modelopt_quant.py) and by the MoE runner hook, and the Muse Spark hook (muse_spark_hook.py) defaults it to True only when unset. It selects per-tensor instead of per-token FP32 activation scales for the target's NVFP4 GEMMs. The draft has no serialized NVFP4 metadata (online config, see above), so it is unaffected. Target precision changes are covered by the passing evals above.
    • 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-20261026-8252e3af@sha256:b4e4abb9… is unchanged by this PR (already on main), comes from the upstream lmsysorg org, and is pinned by digest. The run fingerprint reports SGLang 0.0.0.dev1+g8252e3af6. sgl-project/sglang 8252e3af (#41701) is an ancestor of main (compare 8252e3af...main: ahead 659, 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 GB200 cell musespark-fp4-gb200-sglang-agentic-dflash#0 (Agg TP4·EP4: GPU-resident KV and HiCache items), provenance sha c9eb455f, run 41096285400, pr 4670.
    • 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 41096285400. 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 · gb200_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 41096285400 (attempt 1, head c9eb455f), with x = E2EL (s) and y = request_metrics.throughput.per_gpu.total_tput_tps, computed with infx.workflows.pareto_coverage:
    • P90 E2EL: 7 frontier points / 9 measured (HiCache c7, c10, c14, c16, c28, c36, c40).
    • P75 E2EL (reported separately): 6 frontier points / 9 measured (HiCache c7, c14, c16, c28, c36, c40).
    • No invalid or missing points.
  • Removed points: HiCache c2, c6, c8, c20, c24 and c32. signoff_facts report at c9eb455f finds undominated_removed: [] and unmeasured_removed: [], so every removed point is covered by the measured head curve. Re-tuned deltas (tput/GPU, mean interactivity): c28 −4.3% / +47.5%, c16 −2.6% / +30.1%, c10 −0.3% / +24.1%, c4 −3.7% / +15.1%, c1 +2.2% / +3.7%.
  • Changelog: the entry states the new SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 base env and the point changes. Its claims (prefill-decode interval 64–256 and stream interval 20 at every HiCache point, shortest-prefill-first except at c7, admission 2–6 at c4–c16, 16K chunked prefill at c28–c40, c40 mem 0.93 / HiCache 160 GB / CUDA graphs up to 48) match the resolved per-point configs of the pinned run.
  • Merge commit: c9eb455f merges main into d75601d5. The pinned run is on that merge head itself, so no later commit needs a materiality assessment.
  • CPU DRAM policy: the largest HiCache pool is 160 GB per rank × 4 ranks = 640 GB on one GB200 NVL72 compute tray (4 of 4 GPUs), within the tray's standard CPU DRAM.
  • The changelog pr-link uses the pull/XXX placeholder, which validate_perf_changelog.py accepts (check-changelog passed).
  • 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

github-actions Bot commented Nov 3, 2026

Copy link
Copy Markdown
Contributor

✅✅✅ Verdict: PASS ✅✅✅

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — @nvpohanh is a listed owner of inferencex-e2e/configs/nvidia-master.yaml and of the srt-slurm-recipes/**/gb200-*/ recipe path; perf-changelog.yaml falls only under the catch-all owner.

✅ Check 1 (Sweep on in-PR commit): PASS — Run Sweep 41096285400 (pull_request, attempt 1) ran on the pinned head c9eb455f itself; its canary, all 9 agentic / jobs and all 4 agentic eval / jobs succeeded. There are no later commits to assess.

✅ Check 2 (Evals): PASS — eval_results_all from run 41096285400 has gsm8k em_strict of 0.9795 (c1), 0.9795 (c7), 0.9765 (c16) and 0.9765 (c40), n=1319 each, on the same pinned lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4… image. Every row is within 2σ of its best sibling and of the exact-scope GB200 cohort (run 40279195159, 0.9757). The new base env SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 is in base at the pinned recipe, so it applies to every eval point.

✅ Check 3 (Upstream recipe): PASS — sgl-project/sglang#47929 is MERGED (2026-11-03T03:28:14Z) and the MuseSparkOpen cookbook page is live. Its muse-spark-open.jsx GB200 cell (provenance c9eb455f, run 41096285400) matches every production-useful flag and env var at c1, c4, c7, c10, c14, c16, c28, c36 and c40, including SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0. The exclusions (enable-metrics, PYTHON*, HF_HUB_OFFLINE, harness and benchmark env) are justified in the sign-off.

✅ Check 4 (Reuse command): PASS — nvpohanh (COLLABORATOR) posted /use 41096285400 on its own line, naming the Check 1 run.

✅ Check 5 (Checklist template): PASS — all 20 items in the current PR_REVIEW_CHECKLIST.md template are present and checked.

✅ Check 6 (Upstream image / provenance / ordering): PASS — the image is the unchanged upstream lmsysorg/sglang:nightly-dev-20261026-8252e3af, pinned by digest. SGLang commit 8252e3af is an ancestor of sgl-project/sglang main (ahead 660, behind 0). This is an SGLang (open-source engine) entry, so the engine-first ordering rule does not apply.

✅ Check 7 (Deprecated models): PASS — the musespark agentic-coding scenario is active in MODELS.md, with no deprecation notice as of the 2026-11-03 review date.

✅ Check 8 (Architecture hacks): PASS — there are no --hf-overrides, json-model-override-args or layer/expert skips. The new target per-tensor NVFP4 activation-scale setting is covered by the passing evals in Check 2. musespark is in the policy-baseline matrix, so it is grandfathered for integer all-reduce; it runs on CUDA and sets no *QUICK_REDUCE_QUANTIZATION control.

✅ Check 9 (Chat template): PASS — the AgentX replay drives the chat endpoint (profile_export_aiperf.json in the run 41096285400 c40 artifact shows endpoint "type": "chat") through the shared agentx profile harness, which this PR does not change.

✅ Check 10 (Engine patches): PASS — the diff is YAML only, and the effective recipe has no setup script, patch, pip install or file overwrite.

✅ Check 11 (Golden AL): PASS — every point's submission env has SGLANG_SIMULATE_ACC_LEN=4.05, SGLANG_SIMULATE_ACC_METHOD=match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token. That matches musespark_dflash.yaml thinking_on with 7 draft tokens = 4.05 (signoff_facts golden-al reports matches: true).

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

✅ Check 13 (Draft precision): PASS — the draft is RadixArk/Muse-Spark-Open-DFlash2 (HF sha 46c90f41), stored entirely in BF16 with no quantization_config. At sglang 8252e3af, the inherited modelopt_fp4 resolves to the online config, and nvfp4_online.get_quant_method returns UnquantizedLinearMethod for LinearBase layers. DFlashMLP is dense, with no FusedMoE, so the draft serves in BF16. Non-serialized ModelOptFp4Config forces use_per_token_activation=False, so the new env var does not reach the draft. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE defaults to EnvBool(False) and is not set anywhere.

✅ Check 14 (Pareto coverage): PASS — there is one curve: musespark / agentic-coding / gb200_sglang / fp4 from the run 41096285400 bmk_agentic_* artifacts (head c9eb455f, same image), x = E2EL and y = per_gpu.total_tput_tps. pareto_coverage gives 7/9 frontier points at P90 (c7, c10, c14, c16, c36, c28, c40) and 6/9 at P75 (reported separately), with no invalid points.

✅ Check 15 (PR description): PASS — the per-point changes, removed points, unchanged image and checkpoint, and the base env addition all match the recipe at c9eb455f.

✅ Check 16 (NVIDIA overview): PASS — Muse-Spark-Open has an SGLang checkmark in the Framework Recipes table of the NVIDIA overview.

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

✅ Check 18 (Recipe/image/sweep consistency): PASS — no generation errors at head. The master entry (TP4/EP4, conc [1] plus HiCache [4, 7, 10, 14, 16, 28, 36, 40]) matches the recipe's overrides. No new images; no orchestration or dispatch changes. The changelog claims match the resolved recipe. undominated_removed and unmeasured_removed are both empty. Re-tune deltas (tput per GPU / interactivity): c28 −4.3%/+47.5%, c16 −2.6%/+30.1%, c10 −0.3%/+24.1%, c4 −3.7%/+15.1%, c1 +2.2%/+3.7%.

Assessed commit: c9eb455f32e7f026d7424de723756de2e719d868.

@nvpohanh

nvpohanh commented Nov 4, 2026

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (musespark-fp4-gb200-sglang-agentic-dflash) will happen in sgl-project/sglang#48392, in the Agentic Long-Context Deployment section of the Muse Spark Open cookbook page. The config has been live in that section since sgl-project/sglang#47464; sgl-project/sglang#48392 refreshes it with the pin below.

The commands mirror the launch measured in run 41096285400, pinned by this PR's /use comment.

@nvpohanh

nvpohanh commented Nov 4, 2026

Copy link
Copy Markdown
Collaborator

[by Claude Code] sgl-project/sglang#48392 has been merged. This PR's AgentX config (musespark-fp4-gb200-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 41096285400, pinned by this PR's /use comment.

@nvpohanh

nvpohanh commented Nov 5, 2026

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (musespark-fp4-gb200-sglang-agentic-dflash) will happen in sgl-project/sglang#48634, in the Agentic Long-Context Deployment section of the Muse Spark Open cookbook page. The config has been live in that section since sgl-project/sglang#47464; sgl-project/sglang#48634 refreshes it with the pin below.

The commands mirror the launch measured in run 41096285400, pinned by this PR's /use comment.

@nvpohanh

nvpohanh commented Nov 5, 2026

Copy link
Copy Markdown
Collaborator

[by Claude Code] sgl-project/sglang#48634 has been merged. This PR's AgentX config (musespark-fp4-gb200-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 41096285400, pinned by this PR's /use comment.

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
@nvpohanh
nvpohanh merged commit 9b5f27a into main Nov 6, 2026
30 checks passed
@nvpohanh
nvpohanh deleted the musespark-fp4-gb200-sglang-agentic-retune-1029 branch November 6, 2026 01:45
@github-actions

github-actions Bot commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category new duplicate skipped unmapped failed
benchmark 9 0 9 0 0
eval 4 4 4 0 0
trace 9 0 0 0 0
power 9 0 0 0 0

benchmarkFiles 10, rawBenchmarkRows 18, coveredByPerJobFiles 9, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 5276, gpuMetricSamplesNew 189824, powerPublicationPoints 9, 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。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended)

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

3 participants