Repository navigation
[AgentX] MiniMax-M3.1-Flash B200/GB200/B300 SGLang: add conc 1 low-latency configs / [AgentX] MiniMax-M3.1-Flash B200/GB200/B300 SGLang：新增 conc 1 低延迟配置 - #4509
Conversation
…ang AgentX Adds the MSA-V1 MonoKernel TP4 point at conc 1, following the GB300 recipe. 为 MiniMax-M3.1-Flash 的 B200、GB200、B300 SGLang AgentX 新增 conc 1 低延迟配置，沿用 GB300 的 MSA-V1 MonoKernel TP4 设置。
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}
| # Follows the GB300 Low-Latency cell: the MSA-V1 MonoKernel serves batch-size-1 | ||
| # DSpark steps; larger batches take the standard path. |
There was a problem hiding this comment.
🟡 nit (optional): the new recipes add a comment saying where the config came from, which AGENTS.md's "No history" rule bans. sweep:# Follows the GB300 Low-Latency cell The comment above args: says what the recipe was copied from. The rule says derivation belongs in git and perf-changelog. Fix: drop "Follows the GB300 Low-Latency cell:" from each of the three new recipes. Keep only the technical part, for example # MSA-V1 MonoKernel serves batch-size-1 DSpark steps; larger batches take the standard path.
Why this was flagged
The three new recipe files b200-fp4-mtp/agentic-low-latency.yaml:24, b300-fp4-mtp/agentic-low-latency.yaml:24 and gb200-fp4-mtp/agentic-low-latency.yaml:24 each add # Follows the GB300 Low-Latency cell: the MSA-V1 MonoKernel serves batch-size-1. AGENTS.md "Code comments (mandatory)" covers recipes and says: "Do not write comments that describe how the code evolved, where it was derived from, what it replaced". The opening clause describes where the recipe was derived from, not why a value was chosen. Behaviour does not change. The cost is a convention breach that leaves maintainers a stale pointer once the GB300 recipe changes.
Verification: Each new recipe (b200-fp4-mtp, b300-fp4-mtp and gb200-fp4-mtp agentic-low-latency.yaml) adds the same two lines at 24-25, just above args:, starting "# Follows the GB300 Low-Latency cell". Line 65 of AGENTS.md says: "No history. Do not write comments that describe... where it was derived from". That clause is exactly what that rule bans. No runtime behavior is affected.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40194550246 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40194550246 |
Keeps only the technical note, per the no-history comment rule. 只保留技术说明，符合不写来源历史的注释规则。
|
/use 40194550246 |
|
@ispobock staged run 40194550246.
This run remains available across future |
…g-agentx-low-latency
Add a Low-Latency point at conc 1 for the B200, GB200 and B300 SGLang AgentX configs, following the GB300 recipe from #4456: TP4 with the MSA-V1 MegaKernel serving batch-size-1 DSpark steps. Each SKU gets its own
-low-latencyconfig key so the sweep covers only the new conc 1 point.中文
为 B200、GB200、B300 的 SGLang AgentX 配置新增 conc 1 低延迟点，沿用 #4456 中 GB300 的设置：TP4，由 MSA-V1 MegaKernel 处理 batch size 1 的 DSpark 步骤。每个 SKU 使用单独的
-low-latencyconfig key，sweep 只覆盖新增的 conc 1 点。AI model disclosure
Prepared with Claude Code using
claude-opus-5-5. Reviewed by the human author before submission.