Skip to content

Conversation

@NuojCheng
Copy link
Collaborator

@NuojCheng NuojCheng commented Nov 24, 2025

Description

This PR enables check_vma=True for the sparse_matmul shard_map, and remove some unnecessaary psum, which are backward transpose of jax.lax.pvary, see jax doc.

This PR currently only supports pure FSDP + megablox kernel.

Following tests are outdated.

Performance Improvements

Test Case 1: Deepseek-test(small) + ZeRO1 + GA

  • Metric: Step Time
  • Result: Reduced from 1457ms to 1435ms.
  • MFU Increase: ~1.5%
  • Profiles:

Test Case 2: Deepseek3-test Model

Test Case 3: Deepseek2-16b on v5p-32 Cluster

  • Metric: Step Time
  • Result: Reduced from 17045ms to 16808ms.
  • Improvement: ~1.41% reduction in step time.
  • Details:

Test Case 4: Deepseek3-671b on tpu7x-512

Checklist

Before submitting this PR, please make sure (put X in square brackets):

  • I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
  • I have necessary comments in my code, particularly in hard-to-understand areas.
  • I have run end-to-end tests tests and provided workload links above if applicable.
  • I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

@NuojCheng NuojCheng added the draft Draft PR label Nov 24, 2025
@NuojCheng NuojCheng force-pushed the chengnuojin-moe-checkvma branch 3 times, most recently from fe22e79 to 482336b Compare November 25, 2025 00:12
@NuojCheng NuojCheng force-pushed the chengnuojin-moe-checkvma branch from 482336b to f18bd3e Compare December 10, 2025 19:44
@NuojCheng NuojCheng changed the title [Merge after 2579 and 2737] Make check_vma true for moe sparse matmul Set check_vma true for moe sparse matmul Dec 10, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

draft Draft PR

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant