Repository navigation
[AgentX] MiniMax-M3.1-Flash GB300 SGLang: add more TP2 configs / [AgentX] MiniMax-M3.1-Flash GB300 SGLang：补充 TP2 配置 - #4525
Conversation
… at conc 20-36 Adds a separate TP2 key with L2 HiCache (180 GB per rank) at conc 20, 24, 28 and 36. 为 MiniMax-M3.1-Flash GB300 SGLang AgentX 新增单独的 TP2 + L2 HiCache key（每 rank 180 GB），覆盖 conc 20、24、28、36。
ispobock
added
the
full-sweep-fail-fast
|
Thanks for the contribution!
中文感谢你的贡献！
|
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| minimaxm31flash-fp4-gb300-sglang-agentic-dspark-tp2: | ||
| image: lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509bcb729c653528d448bff9850f8977fcad9d12ba56bd6dbe3d7985e87 | ||
| model: MiniMaxAI/MiniMax-M3.1-Flash | ||
| model-prefix: minimaxm31flash | ||
| runner: cluster:gb300-nv | ||
| precision: fp4 | ||
| framework: sglang | ||
| multinode: false | ||
| srt-recipe-dir: minimaxm31flash/sglang/gb300-fp4-mtp | ||
| scenarios: | ||
| agentic-coding: | ||
| - dram-utilization: 0.80 | ||
| search-space: | ||
| # TP2 with L2 HiCache: 180 GB per rank at conc 20-36. | ||
| - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [20, 24, 28, 36], srt-recipe: agentic-tp2.yaml } |
There was a problem hiding this comment.
🟡 (optional) The new TP2 points (conc 20-36) get their own config key and changelog entry. Readers then see a separate result set, not more points on the existing GB300 SGLang TP2 curve (conc 32-44). CONTRIBUTING.md says a PR that only adds points to an existing curve should add them to the existing key and mark the changelog entry append-only: true. That route already sweeps only the new points, so a separate key is not needed to avoid re-running the old ones. Fix: add this TP2 row to minimaxm31flash-fp4-gb300-sglang-agentic-dspark (dram-utilization 0.80 scenario) and put zip_override_tp2_hicache in agentic.yaml as another override. Then list that key in the changelog block with append-only: true.
Why this was flagged
The PR description says these points sit alongside the TP2 points from #4511. Those points live under minimaxm31flash-fp4-gb300-sglang-agentic-dspark (nvidia-master.yaml:8957, conc 32/40/44, same image, same hicache 180 GB, same mem-fraction 0.87). The diff instead adds a new key, minimaxm31flash-fp4-gb300-sglang-agentic-dspark-tp2 (nvidia-master.yaml:8962), and a new non-append-only changelog block (perf-changelog.yaml:11208). The block's stated reason is to avoid re-sweeping the existing points. CONTRIBUTING.md:172-186 describes append-only: true as the mechanism for exactly this case. It runs only the generated set difference and lets InferenceX-app extend the most recent matching curve, rather than show the partial run as a separate curve. With the separate key, conc 20-36 is published as its own config and TP2 data is split across two keys. Folding the new points into the existing key would keep TP2 as one curve.
Verification: Trigger: the PR merges with the new key and a regular (not append-only) changelog block, so its sweep is ingested as its own run. The new changelog block (perf-changelog.yaml:11208-11213) has no append-only. Because the app gets no append-only metadata, it will show this partial TP2-only run as a separate curve instead of extending the existing one.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40296110489 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40296110489 |
ispobock
added
full-sweep-fail-fast
|
/use 40296110489 |
|
@Oseltamivir staged run 40296110489. This run remains available across future |
|
Ingest report — ✅ Results ingested into the Ingest workflow · Merge Ingest run · Benchmark source run
benchmarkFiles 4, rawBenchmarkRows 4, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 0, gpuMetricSamplesNew 45834, powerPublicationPoints 4, 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 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。 |
Add TP2 points with DRAM KV offload at conc 20, 24, 28 and 36 to the existing GB300 SGLang AgentX curve, alongside the TP2 points added in #4511. They use the same image and serve flags as the existing TP2 points, with a host cache of 180 GB per rank; the recipe lives in its own file, so existing points are unchanged.
中文
为现有的 GB300 SGLang AgentX 曲线新增 conc 20、24、28、36 的 TP2 + DRAM KV offload 点，作为 #4511 中 TP2 点的补充。镜像和启动参数与现有 TP2 点相同，每个 rank 的 host cache 为 180 GB；recipe 单独成文件，现有的点不受影响。
AI model disclosure
Prepared with Claude Code using
claude-opus-5-5. Reviewed by the human author before submission.