config(musespark): configure GB200 AgentX settings per concurrency / config(musespark): 按并发配置 GB200 AgentX 参数

[RohitNagraj]

Configure Muse Spark Open GB200 SGLang AgentX server settings per concurrency and add concurrency 20 and 28. Concurrency 4–24 selects breakable prefill CUDA graphs and smaller 8,192-token prefill chunks (the default is 16,384); higher-concurrency variants also set scheduling, memory allocation, and cache handling individually.

The image digest, model checkpoint, TP4/EP4 topology, HiCache size, and DFlash2 draft checkpoint and token count stay unchanged. At concurrency 10–32, removing the explicit draft attention backend selects the image default and can change the draft KV-cache dtype inherited from the target. The cache slot reclamation setting is also enabled for those variants. Cookbook coverage for the new settings remains a merge requirement.

AI model disclosure

• Model/version: GPT-6; the exact runtime model/version identifier could not be verified.
• Role: prepared the configuration changes, checked the generated configuration, and drafted the PR description.

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

按并发配置 Muse Spark Open GB200 SGLang AgentX 服务器参数，并新增并发 20 和 28。并发 4–24 使用可拆分的预填充 CUDA graph 和较小的 8,192 token 预填充分块（默认值为 16,384）；较高并发的配置还分别设置调度、内存分配和缓存处理参数。

镜像摘要、模型检查点、TP4/EP4 拓扑、HiCache 大小，以及 DFlash2 草稿模型检查点和 token 数量保持不变。并发 10–32 移除显式指定的草稿注意力后端，改用镜像默认值，这可能改变草稿 KV cache 从目标模型继承的数据类型。这些配置还启用了缓存槽位回收。合并前仍需补齐新参数的 cookbook 文档。

AI 模型使用说明：使用 GPT-6，无法确认运行环境的精确模型及版本标识。工作内容包括准备配置改动、检查生成的配置，以及起草 PR 说明。

[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]

Code review found no issues

No high-confidence issues detected in this change.

This review covers commit 7f7b3a0, which is no longer the latest commit on this pull request; later commits are not covered by it.

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=39195111790

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=39195111790

[nvpohanh]

[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#45111, 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 (511afb819) as "in progress". A later cookbook PR switches it to the measured run once a run is pinned or this PR merges.

[RohitNagraj]

/use 39195111790

[github-actions]

@RohitNagraj staged run 39195111790.

• Muse-Spark-Open · agentic-traces · FP4 · gb200_sglang · 2026-10-19

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] sgl-project/sglang#45111 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's /use 39195111790 comment came after this cookbook round was synced, so the cookbook shows the recipe at 511afb819 as "in progress". That is the commit run 39195111790 measured, so the commands are already the measured ones; a later cookbook PR marks them verified.

[nvpohanh]

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

• Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
• Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this.
• 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
• 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
• 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, GB200, TP4/EP4, DFlash2, HiCache), agentic-coding only. Changed files: musespark/sglang/gb200-fp4-mtp/agentic.yaml (per-concurrency overrides, new c20/c28), nvidia-master.yaml (conc-list adds 20 and 28), and one appended perf-changelog.yaml entry. Assessed head: 511afb819454a40704f661d34011cbdff6907cc8.
• PR validation: Run Sweep 39195111790 on head 511afb81: all 12 agentic / jobs (c1, 2, 4, 6, 8, 10, 12, 16, 20, 24, 28, 32) and the agentic eval / job succeeded. Fixed-sequence and multi-node lanes were skipped because this PR has no such configs. The run is pinned with /use 39195111790 by @RohitNagraj (COLLABORATOR). Every point used the full semianalysis_cc_traces_weka_062126 corpus (semianalysisai/cc-traces-weka-062126), not the 256k corpus.
• Evals: agentic eval / (c32 recipe, same image digest sha256:8cfc5a72…bb2b0a): gsm8k em_strict = 0.9803 (± 0.0038, n=1319), above the 0.90 default threshold in infx/evals/thresholds.yaml (no musespark override). Artifact: eval_results_all from run 39195111790.
• Chat template: AgentX replay drives POST /v1/chat/completions (3,943 chat requests and 0 raw completions in the c32 server log), with the server using chat_template=/model/chat_template.jinja.
• Draft precision (served as shipped): The draft is RadixArk/Muse-Spark-Open-DFlash2 (HF revision main, sha 46c90f41). It ships as BF16: config.json has dtype: bfloat16 and no quantization config, and safetensors are 4.89B BF16 params. The recipe sets no draft quantization, dtype, or revision override. In the server logs, _speculative_draft_quantization_explicitly_set=False, and the draft loads through the default loader without a quant tag (Load weight end … type=DFlash2DraftModel, mem usage=3.61 GB per GPU at TP4, consistent with BF16). By contrast, the target logs quant=modelopt_fp4, quant_algo=NVFP4. Effective draft weight precision: BF16 = shipped precision.
  • Draft KV cache: at c1–c8 (speculative-draft-attention-backend: fa4 retained), the image forces BF16 draft KV (DFLASH fa4 draft: overriding KV cache dtype float8_e4m3fn -> bfloat16). At c10–c32 the override is removed (null). The pinned image's default backend (trtllm_mha) then inherits the target's FP8 (fp8_e4m3) KV cache for the draft. Both are the framework's default handling, with no draft-specific KV dtype override (speculative_draft_kv_cache_dtype=None). The FP8 case is an upstream-supported KV dtype applied consistently to target and draft, which the checklist explicitly allows.
  • SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE does not appear in the recipe, the master config, the generated sbatch/config, or the c1/c4/c32 server logs. It is not enabled, and DFlash2 is a dense Qwen3-style draft with no NextN MoE.
• Golden AL: The connector (infx/srt_slurm/synthetic_acceptance.py) maps DFlash speculative-num-draft-tokens: 8 to num_speculative_tokens = 7. musespark_dflash.yaml muse-spark-open-816b-a42b-nvfp4 / thinking_on / 7 = 4.05. The c1, c4 and c32 server logs show SGLANG_SIMULATE_ACC_LEN=4.05, SGLANG_SIMULATE_ACC_METHOD=match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token. The recipe hard-codes none of them.
• MODELS.md: musespark (Muse Spark Open 816B-A42B, added 2026-10-15) lists Agentic coding with the SGLang DFlash2 draft as active, with no deprecated scenarios.
• No architecture hacks: no --hf-overrides or json-model-override-args (effective json_model_override_args='{}'), and no layer, expert or indexer skipping. The changes are scheduling, prefill CUDA graph, chunking, memory and cache-slot tuning only.
• Dynamo TRT-LLM metrics: N/A. This is a direct SGLang deployment with no Dynamo TRT-LLM serving path.
• Upstream image / engine order: lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a (upstream lmsysorg org, unchanged by this PR). This is an SGLang submission, so the engine-first ordering is satisfied.
• Upstream recipe: sgl-project/sglang#45111 is MERGED (2026-10-19T07:24:09Z, merge commit af9bb2a3). 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, provenance sha 511afb81, pr 4302).
  • The GB200 entry documents the full python3 -m sglang.launch_server command matching the recipe base: TP4/EP4, modelopt_fp4, flashinfer_trtllm MoE, fp8_e4m3 KV, layernorm-SP/MegaMoE, DFLASH with DFlash2 and 8 draft tokens, fa4 draft attention, swa-full-tokens-ratio 0.1, HiCache 120, muse parsers, and SGLANG_TIMEOUT_KEEP_ALIVE=900.
  • It also documents every per-concurrency override in this PR, including c20 and c28: max-running-requests/cuda-graph-max-bs-decode, the breakable prefill CUDA graph with chunked-prefill-size/max-prefill-tokens 8192, mem-fraction-static, schedule-policy shortest-prefill-first, prefill-decode-interval, removal of the draft attention backend, and the SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK, SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS and PYTORCH_CUDA_ALLOC_CONF env vars.
  • InferenceX-only settings excluded from upstream coverage: enable-metrics (scraped by the AIPerf harness), PYTHONNOUSERSITE/PYTHONUNBUFFERED, the health-check polling, all benchmark.env client settings (TOKENIZER, AIPERF_*, AGENTIC_WARMUP_GRACE_PERIOD, CONC), the cluster-local served-model-name/staged-checkpoint plumbing, and the synthetic-acceptance env injected by the harness.
• No engine patches: the diff adds only YAML overrides. The recipe has no setup scripts, patches, pip install, file overwrites, or monkey-patching (setup: []), so no waiver is needed.
• append-only: true: N/A. The appended changelog entry does not use append-only.
• Pareto coverage: one affected curve: Muse-Spark-Open · agentic-coding · gb200_sglang · FP4. No other config maps to this series. Measured from the 12 bmk_agentic_musespark_tp4_conc* artifacts of run 39195111790 (head 511afb81, same image), x = E2EL (s) and y = throughput.per_gpu.total_tput_tps. Computed with infx.workflows.pareto_coverage:
  • P90 E2EL: 7 frontier points / 12 measured (c12, c10, c16, c20, c24, c28, c32).
  • P75 E2EL (reported separately): 7 frontier points / 12 measured.
  • No invalid or missing points.
• The PR branch currently conflicts with main only in the append-only inferencex-e2e/perf-changelog.yaml tail, which is resolved at 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 the **/gb200-*/ srt-slurm recipe path; perf-changelog.yaml falls only under the catch-all and is covered by any recognized CODEOWNER.

✅ Check 1 (Passing sweep on in-PR commit): PASS — Pinned head 511afb81 (the PR tip, unchanged) has all 12 agentic / jobs (c1–c32, including the new c20/c28) and agentic eval / with conclusion success in run 39195111790 (attempt 1, headSha 511afb8); the fixed-sequence and multi-node lanes were skipped because this PR has no such configs.

✅ Check 2 (Evals): PASS — eval_results_all from run 39195111790: gsm8k em_strict 0.9803 (±0.0038, n=1319) against the 0.90 default threshold (no musespark override in thresholds.yaml); same image lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72…bb2b0a.

✅ Check 3 (Upstream recipe): PASS — sgl-project/sglang#45111 is MERGED (2026-10-19T07:24:09Z, af9bb2a3). Its muse-spark-open.jsx GB200 entry (provenance sha 511afb8, PR 4302) documents the full sglang.launch_server base command (TP4/EP4, modelopt_fp4, flashinfer_trtllm MoE, fp8_e4m3 KV, layernorm-SP/MegaMoE, DFLASH/DFlash2 with 8 draft tokens, fa4 draft attention, HiCache 120, muse parsers, SGLANG_TIMEOUT_KEEP_ALIVE=900). It also documents every per-concurrency c4–c32 override in this PR, including c20/c28 args and env, matched line by line. Only enable-metrics and harness or plumbing settings are excluded, with reasons given in the sign-off.

✅ Check 4 (Reuse command): PASS — /use 39195111790 posted by RohitNagraj (COLLABORATOR) on 2026-10-19; it targets the passing in-PR run.

✅ Check 5 (Latest checklist template): PASS — All 17 items and the recipe sub-item of the current PR_REVIEW_CHECKLIST.md template are present and checked.

✅ Check 6 (Upstream image / engine order): PASS — framework: sglang with upstream lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72… (unchanged); open-source engine entry only, so the ordering rule is satisfied.

✅ Check 7 (Deprecated models): PASS — MODELS.md lists musespark (added 2026-10-15) with Agentic coding (SGLang DFlash2) active and no deprecated scenarios; this PR only changes that agentic-coding config.

✅ Check 8 (Architecture hacks): PASS — No --hf-overrides, json-model-override-args, or model-file edits; the changes are only scheduling, prefill CUDA graph, chunking, memory, draft-attention-backend, and cache-slot settings.

✅ Check 9 (Spec-decode chat template): PASS — AgentX replays through agentx profile (infx/bench/agentic/replay.py) against the chat-completions endpoint; the c32 point recorded 3942 requests, consistent with the sign-off's 3943 chat requests and 0 raw completions.

✅ Check 10 (Engine patches): PASS — The diff adds only YAML overrides; the recipe has no setup scripts, patches, pip install, or file overwrites.

✅ Check 11 (Agentic golden AL): PASS — The harness maps speculative-num-draft-tokens: 8 to 7 draft tokens (synthetic_acceptance.py); musespark_dflash.yaml thinking_on/7 = 4.05; the c32 job log applies SGLANG_SIMULATE_ACC_LEN=4.05, match-expected, and real-draft-token to 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 config.json has dtype: bfloat16 and no quantization_config. The effective c32 launch command (job log) has no draft quantization, dtype, revision, or draft-KV override. The FP8 draft KV at c10–c32 is inherited consistently from the target's fp8_e4m3 (allowed). SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is absent from the recipe, the harness overrides, and the effective server env.

✅ Check 14 (Pareto coverage): PASS — One curve: musespark / agentic-coding / gb200_sglang / fp4, from the 12 bmk_agentic_* artifacts of run 39195111790 (attempt 1, sha 511afb8, same image), total_tput_tps per GPU vs E2EL. P90: 7/12 frontier points (c12, c10, c16, c20, c24, c28, c32). P75, reported separately: 7/12. Counted by hand with the upper_right frontier rule; the sandbox blocked the stdin redirect that pareto_coverage.py needs, so the helper itself did not run. No invalid points.

✅ Check 15 (PR description): PASS — These claims match the pinned recipe: breakable prefill graphs and 8192 chunks at c4–c24; draft attention backend removed and cache-slot reclamation enabled at c10–c32; image, checkpoint, TP4/EP4, HiCache, and DFlash2 unchanged; c20 and c28 added.

✅ Check 16 (NVIDIA overview): PASS — The NVIDIA overview Framework Recipes table lists Muse-Spark-Open with a populated SGLang cell linking the SGLang cookbook MuseSparkOpen page (GB200).

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

Assessed commit: 511afb819454a40704f661d34011cbdff6907cc8.

[Oseltamivir]

lgtm