Commit 792e5d5

authored

[Feature] Add GQA backward kernel with varlen input (#1082)

* [Feature] Add GQA backward kernel with varlen input * [Lint] * [BugFix] Freeze the memory order of all atomic_add operations * [Lint] * [Lint] * [BugFix] Use release order to boost performance

1 parent bb8b3cd commit 792e5d5Copy full SHA for 792e5d5

2 files changed

+792

-2

lines changed

examples/flash_attention
- example_gqa_bwd_tma_reduce_varlen.py
- example_gqa_fwd_varlen.py

2 files changed

+792

-2

lines changed

Comments

(0)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Commit 792e5d5

2 files changed

2 files changed

File tree

2 files changed

2 files changed

0 commit comments