Skip to content

[Gluon] [Triton 3.9 API] Refactor async_copy calls to use async_load for improved performance - #432

Merged
LoserCheems merged 1 commit into
mainfrom
gluon-backend
Oct 23, 2026
Merged

LoserCheems merged 1 commit into
mainfrom
gluon-backend

Conversation

@LoserCheems

Copy link
Copy Markdown
Collaborator

No description provided.

Copilot AI balanced review requested due to automatic review settings October 23, 2026 07:52
@LoserCheems
LoserCheems merged commit 5e5f4d2 into main Oct 23, 2026
1 check passed

Copilot AI 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.

🟡 Changes recommended

The new API is unavailable in Triton 3.6.0, which the package still permits, so compatibility handling or an updated minimum dependency is needed.

1 open issue
What changed in this PR

This PR migrates the Gluon SM80 forward attention kernel to Triton’s newer asynchronous load API, aiming to improve performance.

Changes:

  • Replaces nine async_copy_global_to_shared calls with async_load.
  • Preserves existing load arguments, masking, and synchronization.
File Description
flash_sparse_attn/​ops/​gluon/​flash_fwd_sm80.py Migrates asynchronous query, key, and value loads to the newer API.

Review effort: Balanced

config, config.copy_row_m_offs(), config.copy_col_k_offs()
)
async_copy.async_copy_global_to_shared(
async_copy.async_load(
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