Skip to content

Conversation

jlamypoirier
Copy link
Collaborator

@jlamypoirier jlamypoirier commented Jun 17, 2025

✨ Description

Fix: #303

  • Set triton version to 3.1.0 since later ones have bugs that break dropless MOE.
  • Add tests for sparse matmul. They don't actually catch the bug in 3.3.0 and 3.3.1, which must involve a more specific corner case (ex. for specific dimensions, stride, dtype).
  • Skip unnecessary triton autotune in tests. (Ex. added tests go from ~5s to ~0.5s)
  • Move functional tests to their own directory.

🔍 Type of change

Select all that apply:

  • 🐛 Bug fix (non-breaking change that addresses a specific issue)
  • 🚀 New feature (non-breaking change that adds functionality)
  • ⚠️ Breaking change (a change that could affect existing functionality)
  • 📈 Performance improvement/optimization (improves speed, memory usage, or efficiency)
  • 🛠️ Code refactor (non-functional changes that improve code readability, structure, etc.)
  • 📦 Dependency bump (updates dependencies, including Dockerfile or package changes)
  • 📝 Documentation change (updates documentation, including new content or typo fixes)
  • 🔧 Infrastructure/Build change (affects build process, CI/CD, or dependencies)

@jlamypoirier jlamypoirier changed the title Fix dropless MOE, add tests for sparse linear Fix dropless MOE, add tests for sparse linear, Jun 17, 2025
@jlamypoirier jlamypoirier marked this pull request as ready for review June 17, 2025 20:56
@jlamypoirier jlamypoirier mentioned this pull request Jun 18, 2025
8 tasks
Base automatically changed from test_all_models to main June 19, 2025 20:40
@jlamypoirier jlamypoirier merged commit 0f77750 into main Jun 19, 2025
0 of 2 checks passed
@jlamypoirier jlamypoirier deleted the fix_dropless_mlp branch June 19, 2025 20:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[bug] MOE is broken with triton 3.2.0, 3.3.0 and 3.3.1

2 participants