MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX: tuned config-only curve on nightly-ac68c308 / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-ac68c308 的纯配置调优曲线 by hshrivastava-droid · Pull Request #4357 · SemiAnalysisAI/InferenceX · GitHub
Skip to content

MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX: tuned config-only curve on nightly-ac68c308 / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-ac68c308 的纯配置调优曲线 - #4357

Open
hshrivastava-droid wants to merge 1 commit into
mainfrom
minimaxm3-b300-fp4-vllm-agentic-config
Open

hshrivastava-droid wants to merge 1 commit into
mainfrom
minimaxm3-b300-fp4-vllm-agentic-config

Conversation

@hshrivastava-droid

Copy link
Copy Markdown
Collaborator

Description

Config-only update for minimaxm3-fp4-b300-vllm-agentic-mtp.

  • Config: minimaxm3-fp4-b300-vllm-agentic-mtp
  • Image: vllm/vllm-openai:nightly-ac68c3087215e0a4f3cdfa218508c6aada57235d@sha256:98adb2311118263a453ec0acd5d4b72bda3db26cecc0fe8ffc94b9ab03c951f9

Type of Change

  • Bug fix
  • New feature
  • Configuration change
  • Documentation update
  • Other (please describe)

Checklist

  • I have tested my changes locally
  • I have updated documentation if necessary
  • If I changed a container image or config, I have already updated perf-changelog.yaml
    • New perf-changelog.yaml entries are appended to the end of the file (the file is chronological: oldest at top, newest at bottom)
  • Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /reuse-sweep-run on this PR — do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps (remove and re-add the label to force one)

@hshrivastava-droid
hshrivastava-droid requested a review from a team October 20, 2026 20:15
@hshrivastava-droid hshrivastava-droid added the full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures label Oct 20, 2026
…htly-ac68c308 / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-ac68c308 的纯配置调优曲线
@hshrivastava-droid
hshrivastava-droid force-pushed the minimaxm3-b300-fp4-vllm-agentic-config branch from 1a6f2c7 to b7c0601 Compare October 20, 2026 20:16

@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.

I reviewed this PR and didn't find any bugs. Both the recipe and the master config are covered by NVIDIA CODEOWNERS, and the tuning picks are judgment calls, so it's worth having an owner take a look too.

What was reviewed:

  • Recipe override names against the new conc-lists (TP8 c1, TP4 c1/c24, TP4 DRAM c36/c40). Each point has exactly one override, and no overrides are left over after the TP2 and c48/c64 removals.
  • The new perf-changelog entry is appended at the end and leaves existing bytes untouched. It names the config key, and the image tag matches the master config.
  • I looked at max-num-seqs: 7 on the TP4 c24 point, which is far below CONC 24 and well under the ~1.5x ratio the other points use. The hunt ruled it out as a bug, but an owner should confirm it's the intended tuning. max-num-active-seqs (c36/c40) is also used by no other recipe in the repo, so it depends on the new nightly image supporting it.
Extended reasoning...

The change is config-only: it bumps the digest-pinned vLLM nightly image for minimaxm3-fp4-b300-vllm-agentic-mtp, shrinks the search space, retunes the matching srt-slurm AgentX recipe overrides and appends one perf-changelog entry. It touches no security-sensitive surface. I deferred instead of approving because nvidia-master.yaml and the b300-* recipe directories have dedicated CODEOWNERS. The new values (max-num-seqs below CONC at c24, and a max-num-active-seqs arg no other recipe uses) are tuning choices that the owners are better placed to confirm.

This review covers commit 1a6f2c7, which is no longer the latest commit on this pull request; later commits are not covered by it.

@github-actions

Copy link
Copy Markdown
Contributor

This branch has not been deployed

No deployments
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: No status

Development

Successfully merging this pull request may close these issues.

1 participant