[AgentX] GB300 DSV4.1 Flash SGLang: 20261012 nightly, 7-point curve with DEP2 and TP4 EP1 / [AgentX] GB300 DSV4.1 Flash SGLang：升级 20261012 nightly，精简为 7 个点（DEP2、TP4 EP1） - #3859

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.

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38341219914

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38341219914

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (dsv41flash-fp4-gb300-sglang-agentic-dspark) will happen in sgl-project/sglang#43799, in the Agentic Long-Context Deployment section of the DeepSeek-V4.1 cookbook page.

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (e182134fd) as "in progress". It switches to the measured run at the next cookbook sync after a run is pinned or this PR merges.

[by Claude Code] sgl-project/sglang#43799 has been merged. This PR's AgentX config (dsv41flash-fp4-gb300-sglang-agentic-dspark) is now in the Agentic Long-Context Deployment section of the SGLang DeepSeek-V4.1 cookbook: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4_1#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 (e182134fd) as "in progress". It switches to the measured run at the next cookbook sync after a run is pinned or this PR merges.

/use 38221097558

@nvpohanh staged run 38221097558.

• DeepSeek-V4.1-Flash · agentic-traces · FP4 · gb300_sglang · 2026-10-12

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

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (dsv41flash-fp4-gb300-sglang-agentic-dspark) will happen in sgl-project/sglang#43983, in the Agentic Long-Context Deployment section of the DeepSeek-V4.1 cookbook page. The config has been live in that section since sgl-project/sglang#43799; sgl-project/sglang#43983 refreshes it with the pin below.

The commands mirror the launch measured in run 38221097558, pinned by this PR's /use comment.

[by Claude Code] sgl-project/sglang#43983 has been merged. This PR's AgentX config (dsv41flash-fp4-gb300-sglang-agentic-dspark) is now in the Agentic Long-Context Deployment section of the SGLang DeepSeek-V4.1 cookbook: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4_1#agentic-long-context-deployment

The commands mirror the launch measured in run 38221097558, pinned by this PR's /use comment.

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
• 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:

• PR validation: Run Sweep 38221097558 ran on the current head 6b07b8e6930edcb8309cb59ca758fce5a3ef3f6b, which is in this PR's commit list. Attempt 2 concluded success. The run is pinned with /use 38221097558, and its artifacts are unexpired (they expire 2027-01-10/11).
  • All 7 dsv41flash-fp4-gb300-sglang-agentic-dspark benchmark jobs succeeded: canary TP4 c1, TP4 c4/c16, TP2/EP2 c16/c32 and DEP2 c64/c128. Every bmk_agentic_* artifact has records_error_dropped: 0; the only dropped records are warmup.
  • In attempt 1, TP4 c16 failed when AIPerf hit an HF Hub 429 Too Many Requests while loading the trace dataset, before any request was sent. That is an infrastructure failure, and fail-fast cancelled the other five jobs. Attempt 2 reran those six. The canary and both evals carried over from attempt 1.
  • The skipped jobs are the empty fixed-sequence and multi-node lanes.
• Evals: gsm8k em_strict was 0.9735 (n=1319) at both DEP2 c128 and TP2/EP2 c32, from eval_results_all.
  • Both clear the default 0.90 threshold in infx/evals/thresholds.yaml, which has no dsv41flash override.
  • Both evals ran on the same image digest with real DSpark verification: no SGLANG_SIMULATE_ACC_* is set in the logged eval worker environment.
  • The equal scores are a coincidence. Only 100 of the 1319 responses are identical between the two runs.
• Image and engine order: upstream lmsysorg/sglang:nightly-dev-cu13-20261012-e6072935@sha256:b5c488694d74c13587af6057b90220498c4869571c8e1e3c43dd6f43ae721306. Every worker fingerprint reports sglang 0.0.0.dev1+ge6072935b and sglang-router 0.3.2, and the master config and recipe model.container match. The config is framework: sglang, so the engine-ordering rule is satisfied.
• Chat template: AgentX replay runs aiperf profile --scenario agentx. The profile_export_aiperf.json of all 7 points records endpoint.type = chat with streaming.
• Draft precision:
  • Draft: the DSpark head bundled in deepseek-ai/DeepSeek-V4.1-Flash (mtp.0–mtp.2, block size 5). All 7 benchmark workers and both eval workers load snapshot 5254448103b6254fc646dacbca1cbbf20e9216c1, with speculative_draft_model_path=deepseek-ai/DeepSeek-V4.1-Flash.
    • This revision differs from 2cba9e42 (the revision used in [AgentX] GB300 DSV4.1 Flash SGLang: fold DSpark sampling, retune c64/c128 / [AgentX] GB300 DSV4.1 Flash SGLang：折叠 DSpark 采样并重调 c64/c128 #3782) only in inference/model.py, inference/requirements.txt and .gitignore. Every safetensors shard and config.json has the same LFS oid.
  • Shipped precision, from the safetensors headers of shards 44–46:
    • attention (wq_a, wq_b, wkv, wo_a, wo_b), shared experts and main_proj are F8_E4M3 with F8_E8M0 scales;
    • the 384 routed experts per stage are MXFP4 (I8-packed with F8_E8M0 scales);
    • norms, gate, markov_head and confidence_head are BF16, and hc_*, attn_sink and the gate biases are F32;
    • quantization_config is fp8 with dynamic activations, 32x32 blocks, ue8m0 scales and expert_dtype: fp4, with no exclusions.
  • Effective precision: the same, through the pinned image's default handling.
    • Every point logs quantization=None, speculative_draft_model_quantization=None, speculative_draft_kv_cache_dtype=None and json_model_override_args='{}'. The KV cache is the default fp8_e4m3 for both target and draft.
    • At e6072935, DeepseekV4DSpark.load_weights always streams wo_a from FP8 to BF16 (_dequant_fp8_wo_a_streaming), which matches DeepSeek's reference (wo_a in dtype=torch.bfloat16).
    • Routed experts, TP4 and TP2: flashinfer_mxfp4 with flashinfer_mxfp4_moe_precision=default, which means MXFP4 weights with MXFP8 activations.
    • Routed experts, DEP2: moe_a2a_backend=megamoe. The draft inherits it (DSpark with DP attention must match the target). Because enable_w4a4_mxfp4_megamoe=False and speculative_enable_w4a4_mxfp4_megamoe=None, MegaMoE uses its default fp8xfp4 MMA. No path uses W4A4.
    • Both paths give FP4 expert weights 8-bit activations, as DeepSeek's reference linear() does (act_quant to FP8, then fp4_gemm).
  • SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: not set in the recipe, the master config, the launch path or the logged worker environment of any point. It defaults to EnvBool(False) in environ.py at e6072935.
  • Other draft-related settings:
    • SGLANG_DSPARK_OPT_MARKOV_W2_BF16=True restates the default EnvBool(True).
    • SGLANG_DSPARK_FOLDED_SAMPLING=2 is deprecated and ignored at e6072935, and the workers log the deprecation warning. Folding the draft proposal into the CUDA graph is the image default (SGLANG_DSPARK_FOLDED_PROPOSAL=True). It affects graph capture only, not weight or activation precision.
• Golden AL: infx/golden_al_distribution/dsv41flash_dspark.yaml gives 3.51 for thinking_on at draft length 5. That matches --speculative-dspark-block-size 5 and SGLANG_DEFAULT_THINKING=1.
  • All 7 benchmark points logged SGLANG_SIMULATE_ACC_LEN=3.51, SGLANG_SIMULATE_ACC_METHOD=match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token.
• MODELS.md: dsv41flash agentic coding is active, with SGLang DSpark arms per SKU, as of 2026-10-13.
• Architecture, and decoder SWA bounded replay (disclosure):
  • There is no --hf-overrides, and json_model_override_args='{}'.
  • TP2/EP2 c16 and c32 enable SGLang's opt-in --enable-decoder-swa-bounded-replay.
    • During prefill, layers 21–39 (the layers after the last kv_source layer, 20) run only over each request's last 128 extend tokens, with their sliding windows starting at the beginning of that tail. Main and indexer KV stay exact.
    • The SGLang DeepSeek-V4.1 cookbook states that it "is not numerically equivalent to full prefill".
    • SGLang rejects the flag with DP attention or the prefill CUDA graph, so the DEP2 and TP4 points do not use it. Their logged enable_decoder_swa_bounded_replay is False.
    • The TP2/EP2 c32 eval ran with the flag and scored 0.9735 on gsm8k.
    • The same flag is in the merged H200 SGLang recipe (#3392, which states the path is not numerically identical) and the merged MI355X SGLang recipe (#3696). ATOM enables the same approximation by default (#3901).
    • Per SA admin guidance, decoder SWA bounded replay is allowed: DeepSeek endorses it, and the model was trained to tolerate it. @cquil11, please confirm.
  • The other changed settings are scheduling, memory or placement settings: the Engram host table location, swa-prefix-tails (SWA pool headroom), prefill-decode-interval, chunk size, admission, mem-fraction-static, tokenizer workers, schedule-policy, load-balance-method, the DP-attention control broadcast, cuda-graph-backend-prefill disabled, SGLANG_NUMA_BIND_V2=0 and expandable_segments. SGLANG_RAGGED_VERIFY_MODE=static restates the pinned default.
  • publish_events_and_metrics appears only in TRT-LLM recipes and does not apply to this SGLang path. --enable-metrics is on.
• No engine patches: the diff changes only the recipe, the master config and the changelog. There is no setup_script, patch file, git apply, heredoc rewrite, site-packages edit or wheel install, and the pinned image runs as shipped. No waiver is needed.
• Recipe publication:
  • Upstream PR: sgl-project/sglang#43983 is MERGED (mergedAt: 2026-10-13T06:39:46Z, merge commit dbad07a06fbedfc63fd54282d2989ef4feb03c23). docs/src/snippets/agentx/deepseek-ai/deepseek-v4_1.jsx is unchanged on SGLang main (d8f85223). The config was first added by sgl-project/sglang#43799, merged 2026-10-12.
  • Published page: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4_1#agentic-long-context-deployment (hardware GB300, KV offload "None", router "SGLang").
  • Variants: it lists the three GB300 deployments in this PR: "Agg TP4" (C1/C4/C16), "Agg TP2·EP2" (C16/C32) and "Agg DEP2" (C64/C128). Each records this PR's head 6b07b8e6, run 38221097558 and the /use comment as its source.
  • Comparison: I decoded all 7 cookbook commands with the cookbook's own decoder (_agentx.jsx). I compared them with the recipe at 6b07b8e6 (base merged with each override, anchors resolved) and with the Env:/Command: lines that srt-slurm logged for all 7 points. All three agree on every worker flag and environment variable, except the InferenceX-only settings below.
    • The DEP2 router command also matches: sglang_router.launch_router --policy cache_aware --request-id-headers x-correlation-id --dp-aware --connect-timeout-secs 900 --request-timeout-secs 14400 --disable-health-check --disable-retries.
    • The pinned image resolves quantization (from the checkpoint config), the KV-cache dtype (fp8_e4m3) and the kernel backends (attention_backend=dsv4, flashinfer_mxfp4 or MegaMoE) itself. Neither the recipe nor the cookbook sets them.
  • Excluded as InferenceX-only:
    • --enable-metrics: the Prometheus export that AIPerf scrapes for result collection.
    • --host, --port, --nccl-port, the router --worker-urls/--prometheus-*, CUDA_VISIBLE_DEVICES and HEAD_NODE_IP: process placement.
    • PYTHONNOUSERSITE and PYTHONUNBUFFERED: Python path isolation and log buffering inside the container.
    • SGLANG_CUDA_COREDUMP_BEFORE_CRASH=0 and SGLANG_PYSPY_DUMP_BEFORE_CRASH=0: crash diagnostics.
    • SGLANG_SIMULATE_ACC_*: the synthetic acceptance control.
    • The AIPerf client settings (CONC, AGENTIC_WARMUP_GRACE_PERIOD, AIPERF_*). This includes AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID, which only makes the benchmark client send the x-correlation-id header that the router's --request-id-headers reads.
• append-only: N/A. The single new changelog entry does not use append-only: true and only appends at the end of perf-changelog.yaml. The current merge conflict with main is limited to that file.
• Code quality:
  • The recipe shares the TP2 and DEP2 settings through YAML anchors, and the header comment describes the new points.
  • The master-config search space matches the overrides: TP2/EP2 c16/c32, TP2/EP2 with dp-attn c64/c128 behind sglang-router 0.3.2, and TP4/EP1 c1/c4/c16.
  • A non-blocking follow-up: the unchanged base env still sets SGLANG_DSPARK_FOLDED_SAMPLING=2, which is now a no-op.
• NVIDIA overview (advisory): DeepSeek-V4.1-Flash is not yet listed in either table of the NVIDIA recipes overview. Only DeepSeek-V4, V4-Flash and V4-Pro are listed.
• Pareto: one curve, dsv41flash / agentic-coding / gb300_sglang / fp4 / run 38221097558. It comes from the 7 bmk_agentic_* artifacts, all on the same image digest (y = per_gpu.total_tput_tps, x = E2EL), and was counted with infx.workflows.pareto_coverage.
  • P90 E2EL frontier: 5 of 7 points (TP4 c16, TP2/EP2 c16, TP2/EP2 c32, DEP2 c64 and DEP2 c128).
  • P75 E2EL frontier: 6 of 7 points, the P90 set plus TP4 c4.
  • Staged dashboard: DeepSeek-V4.1-Flash · agentic-traces · FP4 · gb300_sglang · run 38221097558

Signed: nvpohanh

⚠️ Verdict: WARN ⚠️

❗❗ WARNING ❗❗ NVIDIA official recipes overview: model/framework coverage is missing or unverified. See Check 16.

⚠️ Check 16 (NVIDIA overview): WARN — DeepSeek-V4.1-Flash / SGLang on GB300 has merged upstream coverage in the SGLang DeepSeek-V4.1 cookbook (sgl-project/sglang#43983). It is missing from both tables of the NVIDIA recipes overview, which list only DeepSeek-V4 (SGLang), V4-Flash and V4-Pro. Please add DeepSeek-V4.1-Flash with SGLang to the overview.

Assessed commit: 6b07b8e6930edcb8309cb59ca758fce5a3ef3f6b.

⚠️ Verdict: WARN ⚠️

❗❗ WARNING ❗❗ NVIDIA official recipes overview: model/framework coverage is missing or unverified. See Check 16.

⚠️ Check 16 (NVIDIA overview): WARN — DeepSeek-V4.1-Flash / SGLang on GB300 has merged upstream coverage in the SGLang DeepSeek-V4.1 cookbook (sgl-project/sglang#43983). It is missing from both tables of the NVIDIA recipes overview, which list only DeepSeek-V4 (SGLang), V4-Flash and V4-Pro. Please add DeepSeek-V4.1-Flash with SGLang to the overview.

Passed and not applicable checks
✅ Check 0 (CODEOWNER): PASS — @nvpohanh is a named owner of inferencex-e2e/configs/nvidia-master.yaml and of the **/gb300-*/ srt-slurm recipe path. perf-changelog.yaml falls under the catch-all, which any CODEOWNER satisfies.

✅ Check 1 (Passing sweep on in-PR commit): PASS — Run 38221097558 ran on head 6b07b8e6, which is in the PR. The canary, all 6 agentic / jobs and both agentic eval / jobs are success. This config is agentic-only, so the skipped single-node */, eval / and multi-node lanes have no matrix entries.

✅ Check 2 (Evals): PASS — In eval_results_all, gsm8k em_strict is 0.9735 (n=1319) at both DEP2 c128 and TP2/EP2 c32, above the 0.90 default. Both evals ran on the PR's image lmsysorg/sglang:nightly-dev-cu13-20261012-e6072935@sha256:b5c48869….

✅ Check 3 (Upstream recipe): PASS — sgl-project/sglang#43983 is MERGED (2026-10-13T06:39:46Z). deepseek-v4_1.jsx on SGLang main carries "Agg TP4", "Agg TP2·EP2" and "Agg DEP2" GB300 entries pinned to 6b07b8e6 / run 38221097558, and the published page shows them. Spot-checked overrides match the recipe: TP4 c1/c4/c16, TP2 c16/c32 (16K chunks, bounded replay, prefill graph disabled, interval 96/48, expandable_segments, NUMA_BIND_V2) and DEP2 c64/c128 (DP attention, DP LM head, MegaMoE, total_requests, shortest-prefill-first, interval 24/16). The DEP2 sglang_router command also matches. The InferenceX-only exclusions are justified in the sign-off.

✅ Check 4 (Reuse command): PASS — /use 38221097558 was posted by nvpohanh (COLLABORATOR) in comment 6137670512.

✅ Check 5 (Latest template): PASS — Every item in the current PR_REVIEW_CHECKLIST.md template is present and checked, including the upstream sub-item and the Pareto item.

✅ Check 6 (Upstream image / engine order): PASS — The entry is framework: sglang and uses the upstream lmsysorg/sglang image, digest-pinned, on GB300. This is an open-source engine entry, so the ordering rule does not apply.

✅ Check 7 (Deprecated models): PASS — MODELS.md lists dsv41flash agentic coding as active, with SGLang DSpark arms per SKU, as of 2026-10-13.

✅ Check 8 (Architecture / event publication): PASS — There are no --hf-overrides, override args or model-file edits.

• The TP2/EP2 points use --enable-decoder-swa-bounded-replay. This is the approximation described in DeepSeek's V4.1 report, and it is on by default upstream in vLLM (vllm#56227) and ATOM (feat(v41): decoder SWA bounded replay for prefill (--enable-decoder-swa-bounded-replay) ROCm/ATOM#2507/Streamline agent documentation / 精简智能体文档 #2545). It is already in the merged H200 and MI355X SGLang recipes, and the c32 eval ran with it on (0.9735). The sign-off discloses it.
• publish_events_and_metrics is a TRT-LLM setting that this SGLang path does not support. --enable-metrics is on.

✅ Check 9 (Chat template): PASS — AgentX replay drives the chat-completions endpoint. The sign-off records endpoint.type = chat for all 7 points.

✅ Check 10 (Engine patches): PASS — The diff touches only the recipe, the master config and the changelog. There is no setup script, patch, heredoc or wheel install. The rendered srtctl overrides add only the synthetic-acceptance env.

✅ Check 11 (Golden AL): PASS — dsv41flash_dspark.yaml gives 3.51 for thinking_on at length 5, which matches block size 5 and SGLANG_DEFAULT_THINKING=1. The connector injected SGLANG_SIMULATE_ACC_LEN=3.51 with match-expected / real-draft-token, as seen in the job log and srt-single-node-submission.json.

➖ Check 12 (Append-only): N/A — The new changelog entry has no append-only: true.

✅ Check 13 (Draft as shipped): PASS — The draft is the DSpark head bundled in deepseek-ai/DeepSeek-V4.1-Flash.

• The recipe, overrides and launch path set no draft quantization, dtype or path override.
• SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is unset everywhere and defaults to EnvBool(False) in environ.py at e6072935.
• SGLANG_DSPARK_OPT_MARKOV_W2_BF16=True restates the default.
• The MegaMoE default (fp8xfp4) and the flashinfer_mxfp4 default both keep MXFP4 weights with 8-bit activations, as shipped.

✅ Check 14 (Pareto coverage): PASS — There is one curve: dsv41flash / agentic-coding / gb300_sglang / fp4 / run 38221097558, 7 bmk_agentic_* artifacts on the same image, y = per_gpu.total_tput_tps. pareto_coverage counts 5 frontier points at P90 E2EL (TP4 c16, TP2 c16, TP2 c32, DEP2 c64, DEP2 c128) and 6 at P75 E2EL. No points are invalid.

✅ Check 15 (PR description): PASS — The description's claims match the code at 6b07b8e6: the image, the 7-point search space, TP4 c4/c16 on EP1, the TP2 c16/c32 settings and the DEP2 c64/c128 settings (intervals 24/16, mem-fraction 0.85, admits 256). The local results are clearly labeled as local.

Assessed commit: 6b07b8e6930edcb8309cb59ca758fce5a3ef3f6b.

I am already working on it.

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (dsv41flash-fp4-gb300-sglang-agentic-dspark) will happen in sgl-project/sglang#44065, in the Agentic Long-Context Deployment section of the DeepSeek-V4.1 cookbook page. The config has been live in that section since sgl-project/sglang#43983; sgl-project/sglang#44065 refreshes it with the pin below.

The commands mirror the launch measured in run 38221097558, pinned by this PR's /use comment.

/reuse-sweep-run 38221097558