Skip to content

[NVIDIA] Qwen3.5 B300 NVFP4 SGLang AgentX - #5006

Merged
adibarra merged 3 commits into
mainfrom
qwen35-b300-sglang-agentx
Nov 12, 2026
Merged

adibarra merged 3 commits into
mainfrom
qwen35-b300-sglang-agentx

Conversation

@hshrivastava-droid

Copy link
Copy Markdown
Collaborator

Config

qwen3.5-fp4-b300-sglang-agentic-mtp-pareto

Image

lmsysorg/sglang:nightly-dev-20261013-7224fce1

Type of Change

  • Bug fix
  • New feature
  • Configuration change
  • Documentation update
  • Other

Checklist

  • I have completed the AI model disclosure and kept it current
  • perf-changelog.yaml entry appended, no historical entries edited
  • python -m infx.matrix.changed --base origin/main lists the new key

@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。

@hshrivastava-droid hshrivastava-droid added the full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures label Nov 11, 2026

@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.

Code review found no issues

No high-confidence issues detected in this change.

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (qwen3.5-fp4-b300-sglang-agentic-mtp-pareto) will happen in sgl-project/sglang#49787, in the Agentic Long-Context Deployment section of the Qwen3.5 cookbook page.

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (c01d19232) 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#49787 has been merged. This PR's AgentX config (qwen3.5-fp4-b300-sglang-agentic-mtp-pareto) is now in the Agentic Long-Context Deployment section of the SGLang Qwen3.5 cookbook: https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.5#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 (c01d19232) as "in progress". A later cookbook PR switches it to the measured run once a run is pinned or this PR merges.

@github-actions

Copy link
Copy Markdown
Contributor

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (qwen3.5-fp4-b300-sglang-agentic-mtp-pareto) will happen in sgl-project/sglang#50081, in the Agentic Long-Context Deployment section of the Qwen3.5 cookbook page. The config has been live in that section since sgl-project/sglang#49787; sgl-project/sglang#50081 refreshes it with the pin below.

This PR has no authorized /use <run_id> comment yet, so the cookbook shows the recipe at this PR's head (c01d19232) 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#50081 has been merged. This PR's AgentX config (qwen3.5-fp4-b300-sglang-agentic-mtp-pareto) is now in the Agentic Long-Context Deployment section of the SGLang Qwen3.5 cookbook: https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.5#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 (c01d19232) 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

/use 42970945523

@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

@nvpohanh staged run 42970945523.

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

@nvpohanh

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:

Assessed head c01d1923285dd6cebdecb7668f28899a0f614ad4 (the PR's only commit) against the checklist and verifier on main@0c3f1e24ede23985cf087dae4474e5519a44e37b. The PR adds one new config, qwen3.5-fp4-b300-sglang-agentic-mtp-pareto: a single-node aggregated SGLang AgentX config with 14 points.

  • PR validation / sweep: Run Sweep 42970945523 ran on head c01d1923 and concluded success. Canary, all 14 agentic / jobs (TP4 c4/c6/c12/c14/c16, TP4/EP4 c22/c26, TP2/EP2 c16/c20/c24/c32/c40/c48, TP2/EP2 HiCache c64), 3 agentic eval / jobs, check-changelog, collect-results, collect-evals and calc-success-rate all succeeded. The fixed-sequence and multi-node jobs are skipped because they don't apply to this config. Every job ran on lmsysorg/sglang:nightly-dev-20261013-7224fce1. Reuse is pinned by /use 42970945523.

  • Evals: gsm8k em_strict, n_eff 1319, all from run 42970945523:

    • TP2/EP2 c24: 0.9750 ± 0.0043.
    • TP2/EP2 HiCache c64: 0.9750 ± 0.0043.
    • TP4/EP4 c26: 0.9757 ± 0.0042.

    All three clear the qwen3.5 threshold of 0.94 in infx/evals/thresholds.yaml. The sibling gap is 0.0008, under the 2σ threshold of 0.0121. The exact cohort (B300 FP4 SGLang MTP, run 39017709340) scored 0.9666, so this PR scores higher on every row.

  • Chat template: all 14 server logs show the agentic replay only on POST /v1/chat/completions, with no /v1/completions requests.

  • Draft precision: the draft is the native MTP head embedded in nvidia/Qwen3.5-397B-A17B-NVFP4-V2 (mtp_num_hidden_layers: 1).

    • Shipped: BF16. config.json quantization_config.ignore contains mtp* and mtp.layers.0*. hf_quant_config.json (modelopt 0.45, MIXED_PRECISION) has no mtp entry in quantized_layers. The model-00025-of-00026.safetensors header stores all 1,163 mtp.* tensors as BF16, with no weight scales.
    • Effective: BF16 under the image's default handling. The recipe sets no draft quantization, dtype, or draft-path flag. SGLang's inherited speculative_draft_model_quantization: modelopt_fp4 still honors the checkpoint's ignore list. The Qwen3_5ForCausalLMMTP load uses 8.07 GB/rank at TP2 and 4.08 GB/rank at TP4, which is consistent with unquantized weights.
    • SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE appears nowhere in the recipe, the inherited base, or any of the 14 launch environments/logs. At the image commit 7224fce1 it defaults to EnvBool(False), so it is disabled.
  • Golden AL: the config uses NEXTN 6 steps, topk 1, 7 draft tokens, which is 6 speculative tokens. All 14 server logs show SGLANG_SIMULATE_ACC_LEN=4.8, SGLANG_SIMULATE_ACC_METHOD=match-expected and SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token. That matches qwen3.5-397b-a17b-nvfp4__mtp__native__default-default.yaml thinking_on, 6 tokens = 4.80. signoff_facts golden-al (standard verify) reports matches: true.

  • Model / scenario / architecture: qwen3.5 agentic-coding fp4 is active in MODELS.md. There are no --hf-overrides, no model-config edits, and no layer, expert, or indexer skipping. strip-thinking-cache and HiCache only change which prefixes are cached. They don't change model FLOPs, and the c64 point that uses them is covered by its own eval.

  • Integer all-reduce: qwen3.5 is an existing model on the NVIDIA CUDA path, and there is no quick-reduce quantization control. SGLANG_CUSTOM_ALL_REDUCE_V2_MAX_SIZE_KB is a size threshold, not all-reduce quantization.

  • Dynamo TRT-LLM metrics / append-only: N/A. This is a direct SGLang config, not Dynamo TRT-LLM, and the changelog entry does not use append-only.

  • Image / provenance / ordering: lmsysorg/sglang:nightly-dev-20261013-7224fce1 is an upstream lmsysorg nightly image and resolves now (docker manifest inspect OK). Its commit 7224fce1 is on SGLang main (compare main...7224fce1: ahead 0). This is an SGLang-only submission, so the engine-first ordering is satisfied.

  • No engine patches: the diff only touches the recipe, nvidia-master.yaml and perf-changelog.yaml. There are no setup scripts, patches, monkey-patching, or wheel installs.

  • Cookbook: sgl-project/sglang#49787 (merged 2026-11-11) and sgl-project/sglang#50081 (merged 2026-11-12) are both MERGED. The config is published at Qwen3.5 cookbook, Agentic Long-Context Deployment.

    • The cookbook snippet docs/src/snippets/agentx/Qwen/qwen3.5.jsx on sglang main carries all 14 overrides at c01d1923 (cells Agg TP4, Agg TP4·EP4, Agg TP2·EP2), with python3 -m sglang.launch_server commands.
    • A per-point resolution of every cookbook command matches the recipe's flags and environment variables exactly. That includes MTP 6-1-7, NVFP4 / fp8_e4m3 KV, trtllm_mha, flashinfer_trtllm MoE, the cutedsl all-reduce fusion, --max-running-requests / CUDA-graph caps, the mamba settings, --enable-linear-replayssm-spec, the flashinfer linear-attention prefill, --page-size 64, SGLANG_OPT_MAMBA_SKIP_DECODE_LOCK, SGLANG_CUSTOM_ALL_REDUCE_V2_MAX_SIZE_KB, and the c64 HiCache flags (direct / page_first_direct / 299 GB / write_back, --strip-thinking-cache).
    • Excluded as InferenceX-only: --enable-metrics / --enable-cache-report (benchmark server-metric scraping), PYTHONNOUSERSITE (container hygiene), the simulated-acceptance environment variables, and the CONC / KV_OFFLOADING / TOTAL_CPU_DRAM_GB / AIPERF_* benchmark environment.
  • Stream interval: stream-interval: 50 in the base, inherited by all 14 points (≤ 60). Server logs confirm stream_interval: 50.

  • Pareto (agentic-coding, B300 SGLang FP4, p90 E2EL vs total tput/GPU, results_bmk of run 42970945523 attempt 1): 14 measured, 11 on the frontier (pareto_coverage PASS):

    • TP4 c4 (6.58 s, 7,243)
    • TP4 c12 (8.19 s, 19,134)
    • TP4 c16 (8.55 s, 24,887)
    • TP4/EP4 c22 (9.57 s, 32,562)
    • TP2/EP2 c16 (11.18 s, 45,327)
    • TP2/EP2 c20 (11.90 s, 55,569)
    • TP2/EP2 c24 (12.95 s, 66,913)
    • TP2/EP2 c32 (14.83 s, 76,661)
    • TP2/EP2 c40 (18.24 s, 88,571)
    • TP2/EP2 c48 (20.22 s, 100,089)
    • TP2/EP2 HiCache c64 (30.11 s, 122,765)

    TP4 c6, TP4 c14 and TP4/EP4 c26 are dominated.

  • Branch state: the only merge conflict with main is in inferencex-e2e/perf-changelog.yaml, and this PR only appends an entry to it.

Signed: nvpohanh

@github-actions

Copy link
Copy Markdown
Contributor

✅✅✅ 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 srt-slurm-recipes/**/b300-*/; perf-changelog.yaml falls under the catch-all, which any CODEOWNER satisfies.

✅ Check 1 (Sweep on in-PR commit): PASS — The PR's only commit is pinned head c01d1923, and it has 13 agentic / jobs, a canary job and 3 agentic eval / jobs, all success, in run 42970945523 (headSha c01d1923, attempt 1). The fixed-sequence jobs are skipped because they don't apply to this config. No commits follow it, and the head has not moved.

✅ Check 2 (Evals): PASS — eval_results_all from run 42970945523 has gsm8k em_strict of 0.9750 (TP2/EP2 c24), 0.9750 (TP2/EP2 HiCache c64) and 0.9757 (TP4/EP4 c26), all with n_eff 1319. The eval jobs ran on lmsysorg/sglang:nightly-dev-20261013-7224fce1. The sibling gap is 0.0008 (2σ threshold 0.0121). The exact cohort, run 39017709340, scored 0.9666, so every row is at or above it. No conditional backend change needed separate path coverage, and the HiCache and strip-thinking-cache point is evaluated directly.

✅ Check 3 (Upstream recipe): PASS — sglang#49787 and sglang#50081 are both MERGED. docs/src/snippets/agentx/Qwen/qwen3.5.jsx on sglang main carries all 14 overrides at c01d1923, with sglang.launch_server commands for the Agg TP4, TP4·EP4 and TP2·EP2 cells. The flags and env match the recipe: MTP 6-1-7, modelopt_fp4, fp8_e4m3 KV, trtllm_mha, flashinfer_trtllm, the cutedsl all-reduce fusion, the per-point caps and CUDA-graph sizes, the mamba settings, page-size 64, the flashinfer linear-attention prefill, the HiCache flags and SGLANG_* env. The listed exclusions (metrics, PYTHONNOUSERSITE, simulated acceptance and benchmark env) are InferenceX-only.

✅ Check 4 (Reuse command): PASS — /use 42970945523 was posted on its own line by nvpohanh, who has write/write permission, and it names the Check 1 run.

✅ Check 5 (Checklist template): PASS — All 20 items of the current PR_REVIEW_CHECKLIST.md, including the cookbook sub-item, are present and checked.

✅ Check 6 (Upstream images / provenance / ordering): PASS — The image is the upstream lmsysorg/sglang:nightly-dev-20261013-7224fce1. Its commit 7224fce1 is on sglang main (compare main...7224fce1: ahead 0). qwen3.5 is an established model, and the PR adds only an SGLang entry.

✅ Check 7 (Deprecated models): PASS — MODELS.md lists qwen3.5 agentic coding as active for fp8/fp4, and this is an fp4 agentic-coding config.

✅ Check 8 (No architecture hacks): PASS — There are no --hf-overrides, model-config edits, or layer, expert or indexer skipping. HiCache and strip-thinking-cache only change caching, and the c64 point that uses them has its own eval. There are no quick-reduce or integer all-reduce controls on this CUDA path; SGLANG_CUSTOM_ALL_REDUCE_V2_MAX_SIZE_KB is a size threshold, and qwen3.5 is grandfathered anyway.

✅ Check 9 (Chat template for spec decode): PASS — The AgentX replay client drives /v1/chat/completions (agentx-standalone.md), and the sign-off records chat-only requests in all 14 server logs.

✅ Check 10 (No engine patches): PASS — The diff touches only the recipe YAML, nvidia-master.yaml and perf-changelog.yaml. There is no setup script, patch, source edit or wheel install on the reachable path.

✅ Check 11 (Golden AL): PASS — The connector injects SGLANG_SIMULATE_ACC_LEN=4.8, match-expected and real-draft-token, confirmed in the job logs for c64, TP4/EP4 c26 and TP4 c6. signoff_facts golden-al for qwen3.5 MTP, 6 tokens, thinking_on, standard verify gives 4.8 (matches: true).

✅ Check 13 (Draft precision): PASS — The draft is the embedded MTP head of nvidia/Qwen3.5-397B-A17B-NVFP4-V2. Its config.json (modelopt_mixed) ignores mtp* and mtp.layers.0*, so the head ships in BF16. The recipe sets no draft quantization, dtype or path override. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is absent from the recipe and job logs, and it defaults to EnvBool(False) at 7224fce1 environ.py:1057.

✅ Check 14 (Pareto coverage): PASS — For the curve qwen3.5 / agentic-coding / B300 SGLang / fp4 / run 42970945523 / p90 E2EL vs total tput per GPU, results_bmk (attempt 1, image nightly-dev-20261013-7224fce1) has 14 measured points. pareto_coverage finds 11 on the frontier and reports PASS; TP4 c6, TP4 c14 and TP4/EP4 c26 are dominated.

✅ Check 15 (PR description): PASS — The PR body's config key and image match nvidia-master.yaml at c01d1923.

✅ Check 16 (NVIDIA overview): PASS — The NVIDIA overview's Framework Recipes table lists Qwen3.5 with an SGLang link to the Qwen3.5 cookbook, which covers this B300 SGLang pair.

✅ Check 18 (Recipe/image/sweep consistency): PASS — All 14 srt-recipe overrides resolve, with no generation errors. Master tp/ep/conc match each override's TP, EP and GPU count. The image resolves now (docker manifest inspect OK) and is already used by qwen3.5-fp4-b300-dynamo-sglang-agentic-pareto. No orchestration pins or dispatch changes. The changelog description is accurate. No points are removed or re-tuned. stream-interval: 50 (≤ 60) is inherited by all 14 points and confirmed in the logs.

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

➖ Check 17 (Dynamo metrics): N/A — This is a direct SGLang config, not Dynamo TRT-LLM.

Assessed commit: c01d1923285dd6cebdecb7668f28899a0f614ad4.

…entx

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
…entx

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
@adibarra
adibarra merged commit 3529a9e into main Nov 12, 2026
29 checks passed
@adibarra
adibarra deleted the qwen35-b300-sglang-agentx branch November 12, 2026 23:43
@github-actions

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 14 0 14 0 0
eval 3 3 3 0 0
trace 14 0 0 0 0
power 14 0 0 0 0

benchmarkFiles 15, rawBenchmarkRows 28, coveredByPerJobFiles 14, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 3957, gpuMetricSamplesNew 181166, powerPublicationPoints 14, 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-enabled Full sweep with canary gate; matrix jobs run to completion despite failures

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

3 participants