Skip to content

[AgentX] Retune Muse Spark Open B200 SGLang per concurrency / [AgentX] 按并发重新调优 Muse Spark Open B200 SGLang - #4707

Merged
nvpohanh merged 2 commits into
mainfrom
musespark-fp4-b200-sglang-agentic-retune-1101
Nov 6, 2026
Merged

nvpohanh merged 2 commits into
mainfrom
musespark-fp4-b200-sglang-agentic-retune-1101

Conversation

@shyeh25

@shyeh25 shyeh25 commented Nov 1, 2026

Copy link
Copy Markdown
Collaborator

Summary

Move the Muse Spark Open B200 SGLang AgentX curve to lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc and retune it per concurrency:

  • TP8/EP8, GPU-resident KV: concurrency 1, 4, 10
  • TP8/EP8, DRAM HiCache: concurrency 64
  • TP4/EP4, DRAM HiCache: concurrency 7, 16, 24, 40

TP8 concurrency 2, 6, 8 and 56 and TP4 concurrency 8 and 10 are removed. Each removed point is covered by the new curve on throughput per GPU and mean interactivity: TP8 concurrency 2 is beaten by TP8 concurrency 4, TP8 concurrency 6 and 8 by TP8 concurrency 10, TP8 concurrency 56 lies below the line between TP4 concurrency 40 and TP8 concurrency 64, and TP4 concurrency 8 and 10 lie below the line between TP4 concurrency 7 and 16.

Per-point changes:

  • TP8 concurrency 1:
    • stream interval 40
  • TP8 concurrency 4:
    • request admission and decode graph batch 4 → 2
    • prefill-decode interval 64 → 128
    • stream interval 40
  • TP8 concurrency 10:
    • request admission and decode graph batch 10 → 3
    • prefill-decode interval 64 → 128
    • stream interval 40
  • TP8 concurrency 64:
    • request admission and decode graph batch 64 → 56
    • stream interval 20
  • TP4 concurrency 7 (new), HiCache 120 GB:
    • request admission and decode graph batch 2
    • shortest-prefill-first scheduling
    • prefill-decode interval 64
    • stream interval 20
  • TP4 concurrency 16:
    • request admission and decode graph batch 32 → 6
    • prefill-decode interval 48 → 44
    • stream interval 20, incremental streaming output
  • TP4 concurrency 24:
    • stream interval 20
  • TP4 concurrency 40 (new):
    • request admission and decode graph batch 48
    • memory fraction 0.93
    • chunked prefill and max prefill tokens 16384
    • HiCache size 160 GB
    • shortest-prefill-first scheduling
    • prefill-decode interval 64
    • stream interval 20
    • prefill CUDA graph disabled
    • SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=4096, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

The checkpoint and the DFlash2 draft settings are unchanged.

AI model disclosure
  • Model/version: claude-opus-5-5.
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

摘要

将 Muse Spark Open B200 SGLang AgentX 曲线切换到 lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc,并按并发重新调优:

  • TP8/EP8,KV 驻留 GPU:并发 1、4、10
  • TP8/EP8,DRAM HiCache:并发 64
  • TP4/EP4,DRAM HiCache:并发 7、16、24、40

移除 TP8 并发 2、6、8、56 与 TP4 并发 8、10。每个被移除的点在每 GPU 吞吐与平均交互性上都被新曲线覆盖:TP8 并发 2 被 TP8 并发 4 超越,TP8 并发 6、8 被 TP8 并发 10 超越,TP8 并发 56 位于 TP4 并发 40 与 TP8 并发 64 的连线下方,TP4 并发 8、10 位于 TP4 并发 7 与 16 的连线下方。

各点改动:

  • TP8 并发 1:stream interval 40
  • TP8 并发 4:请求准入与 decode graph batch 4 → 2;prefill-decode interval 64 → 128;stream interval 40
  • TP8 并发 10:请求准入与 decode graph batch 10 → 3;prefill-decode interval 64 → 128;stream interval 40
  • TP8 并发 64:请求准入与 decode graph batch 64 → 56;stream interval 20
  • TP4 并发 7(新增,HiCache 120 GB):请求准入与 decode graph batch 2;shortest-prefill-first 调度;prefill-decode interval 64;stream interval 20
  • TP4 并发 16:请求准入与 decode graph batch 32 → 6;prefill-decode interval 48 → 44;stream interval 20,incremental streaming output
  • TP4 并发 24:stream interval 20
  • TP4 并发 40(新增):请求准入与 decode graph batch 48;memory fraction 0.93;chunked prefill 与 max prefill tokens 16384;HiCache 160 GB;shortest-prefill-first 调度;prefill-decode interval 64;stream interval 20;关闭 prefill CUDA graph;SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=4096、PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

checkpoint 与 DFlash2 draft 设置不变。

🤖 Generated with Claude Code

Move to lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc, retune TP8
concurrency 1, 4, 10 and 64 and TP4 concurrency 16 and 24, add TP4
concurrency 7 and 40, and drop TP8 concurrency 2, 6, 8 and 56 and TP4
concurrency 8 and 10.

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

github-actions Bot commented Nov 1, 2026

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 NVIDIA master config and the b200 srt-slurm recipes have designated code owners, and the retuned values are benchmark-tuning choices, so a code owner should still take a look.

What was reviewed:

  • Each conc-list entry in the master config (TP8 1/4/10, TP8 hicache 64, TP4 hicache 7/16/24/40) has a matching override_tp{N}_c{CONC}, and each override's benchmark.env.CONC matches its name. No orphan overrides are left.
  • Several overrides now admit fewer requests than CONC (tp8_c10 admits 3, tp4_c16 admits 6, tp8_c64 admits 56). The PR description explains this as deliberate tuning, and cuda-graph-max-bs-decode matches max-running-requests in each case.
  • The perf-changelog entry is appended only at the end of the file and names the config key, the scenario, the new image and the concurrencies added and dropped. The pull/XXX placeholder follows an existing pattern.
Extended reasoning...

The change switches the image for musespark-fp4-b200-sglang-agentic-dflash in inferencex-e2e/configs/nvidia-master.yaml, rewrites the per-concurrency overrides in the single-node b200-fp4-mtp/agentic.yaml recipe, and appends one perf-changelog entry. It touches no security-sensitive surface and contains no shell. The diff is consistent across all three files. I deferred instead of approving because .github/CODEOWNERS assigns both nvidia-master.yaml and the b200-* recipe directories to specific owners, and the tuning values can only be validated by a benchmark sweep.

@shyeh25 shyeh25 added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Nov 1, 2026
@github-actions

github-actions Bot commented Nov 1, 2026

Copy link
Copy Markdown
Contributor

@nvpohanh

nvpohanh commented Nov 2, 2026

Copy link
Copy Markdown
Collaborator

/use 41297696082

@github-actions

github-actions Bot commented Nov 2, 2026

Copy link
Copy Markdown
Contributor

@nvpohanh staging run 41297696082. 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 integration of this PR's AgentX config (musespark-fp4-b200-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 41297696082, 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-b200-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 41297696082, 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-b200-sglang-agentic-dflash (single-node aggregated SGLang on B200 with DFlash2), agentic-coding only. Changed files: musespark/sglang/b200-fp4-mtp/agentic.yaml, nvidia-master.yaml (image nightly-dev-20261021-c1c287a4 → nightly-dev-cu13-20261030-302b78cc; TP8/EP8 c1, 4, 10, TP8/EP8 HiCache c64 and TP4/EP4 HiCache c7, 16, 24, 40) and one appended perf-changelog.yaml entry. Assessed head: c965ca2e, the only commit in the PR.
  • PR validation: Run Sweep 41297696082 (attempt 1, pull_request) on head c965ca2e. All 8 agentic / jobs succeeded (TP8 c1, 4, 10; TP8 HiCache c64; TP4 HiCache c7, 16, 24, 40), as did all 6 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 41297696082, 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 41297696082, gsm8k em_strict (n=1319): TP8 c1 0.9757, c10 0.9765, HiCache c64 0.9795; TP4 HiCache c16 0.9750, c24 0.9727, c40 0.9773. 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 B200 cohort (run 39756010759, TP8 c64, 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 2.60–2.64 GB per GPU at TP8 and 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: TP8 c1, c4 and c10 use fa4 draft attention, and the image forces BF16 draft KV there (overriding KV cache dtype torch.float8_e4m3fn -> torch.bfloat16). TP8 c64 and every TP4 point 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.
    • 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 B200 (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 B200 cells musespark-fp4-b200-sglang-agentic-dflash#0 (Agg TP8·EP8, GPU-resident KV and HiCache items) and #2 (Agg TP4·EP4, HiCache), provenance sha c965ca2e, run 41297696082, pr 4707.
    • 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 41297696082. 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 · b200_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 41297696082 (attempt 1, head c965ca2e), 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 / 8 measured (TP8 c4, c10; TP4 HiCache c7, c16, c24, c40; TP8 HiCache c64).
    • P75 E2EL (reported separately): 7 frontier points / 8 measured (TP8 c4, c10; TP4 HiCache c7, c16, c24, c40; TP8 HiCache c64).
    • No invalid or missing points.
  • Removed points: TP8 c2, c6, c8, HiCache c56 and TP4 HiCache c8, c10. signoff_facts report at c965ca2e finds undominated_removed: [] and unmeasured_removed: [], so every removed point is covered by the measured head curve. Re-tuned deltas (tput/GPU, mean interactivity): TP8 c4 −1.4% / +27.9%, TP8 c10 −1.4% / +25.9%, TP8 c1 +1.9% / +17.4%, TP4 c16 −0.1% / +12.5%, TP8 c64 +2.0% / +10.9%, TP4 c24 +0.8% / +7.2%.
  • CPU DRAM policy: HiCache uses 160 GB per rank × 8 = 1.28 TB at TP8 c64 (cap 3 TB), and at most 180 GB per rank × 4 = 720 GB at TP4 (4 of 8 GPUs; cap 1.5 TB) on HGX B200.
  • 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 listed on inferencex-e2e/configs/nvidia-master.yaml and srt-slurm-recipes/**/b200-*/; perf-changelog.yaml falls only under the * catch-all.

✅ Check 1 (Sweep + evals on in-PR commit): PASS — PR head c965ca2e (the only commit; head unchanged) has all 8 agentic / and 6 agentic eval / check-runs success in run 41297696082 (attempt 1, pull_request); no later commits.

✅ Check 2 (Evals accuracy): PASS — eval_results_all gsm8k em_strict TP8 c1 0.9757, c10 0.9765, c64 0.9795; TP4 c16 0.9750, c24 0.9727, c40 0.9773 (n=1319), all on lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc; every row is ok against the best sibling and the exact B200 cohort (run 39756010759, 0.9757); no conditional kernel path requiring dispatch proof.

✅ Check 3 (Upstream recipe): PASS — sgl-project/sglang#47929 MERGED 2026-11-03T03:28:14Z; muse-spark-open.jsx on sglang main has musespark-fp4-b200-sglang-agentic-dflash#0/#2 with provenance c965ca2e/run 41297696082 and base+per-point launch args/env matching every recipe override (TP8 c1/c4/c10/c64, TP4 c7/c16/c24/c40); exclusions (--enable-metrics, --enable-cache-report, host/port plumbing, harness SGLANG_SIMULATE_ACC_*) are justified in the sign-off.

✅ Check 4 (Reuse command): PASS — nvpohanh (COLLABORATOR) posted /use 41297696082 on its own line, naming the Check 1 run.

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

✅ Check 6 (Upstream images / main-branch / engine-first): PASS — lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc is upstream; sglang 302b78ccf is an ancestor of main (behind 0); framework: sglang, so the ordering rule is not triggered.

✅ Check 7 (Deprecated models/scenarios): PASS — musespark agentic coding (SGLang, DFlash2) is active in MODELS.md as of 2026-11-03.

✅ Check 8 (No architecture hacks): PASS — no --hf-overrides/json-model-override-args/layer, expert or indexer skipping; no accuracy-trading levers; musespark is in the policy-baseline matrix (grandfathered), and the CUDA path sets no *QUICK_REDUCE_QUANTIZATION.

✅ Check 9 (Spec-decode chat templates): PASS — AgentX replays through AIPerf chat completions (/v1/chat/completions) with the model chat template.

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

✅ Check 11 (Golden AL): PASS — harness injects SGLANG_SIMULATE_ACC_LEN=4.05 with match-expected/real-draft-token (job log, TP4 c40); musespark_dflash.yaml thinking_on, 7 draft tokens = 4.05 (signoff_facts golden-al matches: true).

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

✅ Check 13 (Draft precision): PASS — draft RadixArk/Muse-Spark-Open-DFlash2 (HF sha 46c90f41) ships all-BF16; launch has no draft quant/dtype/KV overrides; at sglang 302b78ccf nvfp4_online.get_quant_method returns UnquantizedLinearMethod for non-FP8 LinearBase and DFlash2 has no FusedMoE, so draft stays BF16; FP8 draft KV on HiCache points is inherited from the FP8 target (allowed); SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is absent from recipe and worker env.

✅ Check 14 (Pareto coverage): PASS — one curve (Muse-Spark-Open · agentic-coding · b200_sglang · FP4, run 41297696082 bmk_agentic_*, tput_per_gpu vs E2EL): P90 7/8 frontier points, P75 7/8, via pareto_coverage.

✅ Check 15 (PR description): PASS — per-point changes, added/removed points, image, and removal-dominance claims match the pinned recipe, master config, and measured results.

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

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

✅ Check 18 (Recipe/image/sweep consistency): PASS — all 8 points generate at head; master TP/EP/conc lists match recipe overrides (TP4 gpus: 4); image resolves; no orchestration pin or dispatch changes; changelog accurate; undominated_removed/unmeasured_removed empty; re-tune deltas tput/intvty: TP8 c1 +1.9%/+17.4%, c4 −1.4%/+27.9%, c10 −1.4%/+25.9%, c64 +2.0%/+10.9%, TP4 c16 −0.1%/+12.5%, c24 +0.8%/+7.2%.

Assessed commit: c965ca2eb81db733ad280ea693025389ac11014b.

@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-b200-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#47929; sgl-project/sglang#48392 refreshes it with the pin below.

The commands mirror the launch measured in run 41297696082, 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-b200-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 41297696082, 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-b200-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#47929; sgl-project/sglang#48634 refreshes it with the pin below.

The commands mirror the launch measured in run 41297696082, 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-b200-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 41297696082, pinned by this PR's /use comment.

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
@nvpohanh
nvpohanh merged commit 3306a05 into main Nov 6, 2026
30 checks passed
@nvpohanh
nvpohanh deleted the musespark-fp4-b200-sglang-agentic-retune-1101 branch November 6, 2026 01:42
@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 8 0 0 0 0
eval 6 6 6 0 0
trace 8 0 0 0 0
power 8 0 0 0 0

benchmarkFiles 8, rawBenchmarkRows 8, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 7914, gpuMetricSamplesNew 250948, powerPublicationPoints 8, 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