Skip to content

Conversation

@reubenconducts
Copy link
Contributor

The purpose of this PR is twofold:

  1. to add support for mma_pv_is_rs == False, which is currently always set to True
  2. to allow 3 mma warp groups, thus enabling m_block_size == 192, and as a result head dimension 96.

Currently, the kernel hangs with m_block_size == 192 due to launching too few threads per block and attempting to allocate too many registers. This fix puts the CuTe DSL kernel in line with the C++ version; see e.g. the register allocation logic here.

@reubenconducts reubenconducts changed the title [FIX] Allow m_block_size == 192 and mma_pv_is_rs == False in CuTe DSL [FIX] Allow m_block_size == 192 and mma_pv_is_rs == False in Sm90 CuTe DSL Sep 2, 2025
@tridao tridao merged commit 6387433 into Dao-AILab:main Sep 2, 2025
@reubenconducts reubenconducts deleted the rstern/num-mma-wg-fix branch September 3, 2025 02:48
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.

2 participants