feat(config): add Qwen3.8 GB300 AgentX / feat(config): 新增 Qwen3.8 GB300 AgentX 配置 - #3816

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=37714835620

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=37714835620

/use 37714835620

@RohitNagraj staged run 37714835620: https://inferencemax-app-git-staging-semianalysisai.vercel.app/inference?i_dates=2026-10-08~r37714835620

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

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:

• Assessed head: e67ca6fc9ecf4cd57f27654bda8a3e9059923820. Policy read at main@255b9b89. New config qwen3.8next-fp4-gb300-sglang-agentic-mtp (single-node agg TP1/EP1, agentic-coding).

• Validation: run 37714835620, attempt 1 ran on the head commit. All 8 agentic / jobs (c1–c64) and the agentic eval / job passed. The authorized /use 37714835620 is already posted.

• Evals: GSM8K em_strict 97.35% (± 0.44%, n=1319). The eval ran on the same digest-pinned image with real MTP verification; no SGLANG_SIMULATE_ACC_* was set for the eval server.

• Chat template: the aiperf agentx client uses the chat endpoint (profile_export_aiperf.json: endpoint.type=chat). The server auto-detects the Qwen3 chat template with thinking on by default.

• Golden AL: infx.srt_slurm.synthetic_acceptance injects the AL automatically from qwen3.8next_mtp.yaml. The config uses thinking_on with speculative-num-steps 3, so the golden AL is 2.32. The runtime environment confirms SGLANG_SIMULATE_ACC_LEN=2.32, match-expected, real-draft-token. The recipe sets no manual AL.

• Draft precision: the embedded MTP head comes from the same checkpoint (nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54; draft revision pinned to the same SHA). In the shipped hf_quant_config.json, mtp.layers.0.mlp.experts is FP8_BLOCK_SCALES (128×128). The rest of the MTP head is BF16; per the model card it matches Qwen/Qwen3.8-Flash-Next-FP8 byte-for-byte. The pinned image's default handling is the modelopt_mixed per-layer loader, which the draft inherits; the log shows type=Qwen4ExpForCausalLMMTP, quant=modelopt_mixed, quant_algo=MIXED_PRECISION. Effective draft precision therefore matches what ships (FP8 block-scaled routed experts, BF16 elsewhere). SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0 is set explicitly and confirmed in the runtime environment. Target and draft KV are both BF16 (speculative_draft_kv_cache_dtype=bfloat16).

• Architecture / metrics: no --hf-overrides, json-model-override-args or layer/expert trimming. publish_events_and_metrics is a TRT-LLM setting and does not apply to SGLang; --enable-metrics and --enable-cache-report are on.

• Image / ordering: lmsysorg/sglang:v0.5.21-cu130@sha256:1f176b3a… is upstream (SGLang build e00930c5). This is an SGLang entry, so engine-first ordering is satisfied (existing qwen3.8next SGLang entries exist on B200/B300/H200).

• MODELS.md: qwen3.8next is active for Agentic coding (added 2026-08-26), with native/upstream SGLang listed as the primary engine.

• Recipe: sgl-project/sglang#43159 is MERGED (2026-10-08T23:56Z, 2b94cb4e). It adds the GB300 long-context section and the GB300 → NVFP4 (NVDA) → Low Latency → Single Node builder cell. That cell carries the same digest, both checkpoint revisions, all five serving env vars and every non-default serving flag. 15 flags are omitted because they equal the pinned build's defaults. I checked each one against sgl-project/sglang@e00930c5:

  • TP/PP/DP/EP default to 1.
  • mamba-track-interval defaults to 256 and max-prefill-tokens to 16384.
  • moe-a2a-backend defaults to none and decode CUDA graphs default to full.
  • hicache-io-backend defaults to kernel, hicache-mem-layout to page_first and ple-offload-backend to pinned.
  • tokenizer-worker-num and scheduler-recv-interval default to 1.
  • served-model-name falls back to the model path, and speculative-moe-runner-backend inherits the target's flashinfer_trtllm.

  Excluded as InferenceX-only: the trace dataset (WEKA_LOADER_OVERRIDE), simulated-acceptance env vars, CONC overrides, TOTAL_CPU_DRAM_GB/KV_OFFLOADING binding checks, and Slurm/srun/health-check settings.

• Engine patches: none. The recipe uses the stock image with no overlay, patch, sed, pip install or waiver.

• Append-only: N/A. The changelog entry does not set append-only; this is a new config.

• Pareto (run 37714835620, total tput/GPU vs E2EL): one curve, qwen3.8next / agentic-coding / gb300_sglang / fp4. Counts come from the upstream pareto_coverage.py helper on the bmk_* artifacts.

  • p90 E2EL: 5/5 frontier points (c4, c8, c16, c32, c48). c1, c2 and c64 are dominated.
  • p75 E2EL: 7/5 frontier points.
  • All 8 measured points are valid.

• Note: the PR currently has a merge conflict with main. Resolve it without dropping e67ca6fc from the PR history, so /use 37714835620 stays valid for reuse.

Signed: Ankur-singh

✅✅✅ Verdict: PASS ✅✅✅

Assessed commit: e67ca6fc9ecf4cd57f27654bda8a3e9059923820.

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (qwen3.8next-fp4-gb300-sglang-agentic-mtp) will happen in sgl-project/sglang#43083, in the Agentic Long-Context Deployment section of the Qwen3.8-Flash-Next cookbook page. It mirrors the launch measured in run 37714835620, pinned by this PR's /use comment.

[by Claude Code] sgl-project/sglang#43083 has been merged. This PR's AgentX config (qwen3.8next-fp4-gb300-sglang-agentic-mtp) is now in the Agentic Long-Context Deployment section of the SGLang Qwen3.8-Flash-Next cookbook: https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-Flash-Next#agentic-long-context-deployment

The commands mirror the launch measured in run 37714835620, pinned by this PR's /use comment.