Fuse per-sequence AlltoAll into a unified one in GDN forward - #4913

Merged

xuantengh merged 13 commits into

NVIDIA:mainfrom

xuantengh:xuantengh/gdn_a2a

Jun 9, 2026

Commits on Jun 9, 2026

refactor a2a op to reuse existing code from mapping.py
xuantengh
committed
2d66c4e
fuse per-seq a2a into a unified one
xuantengh
committed
7d8f381
use qwen3 model config for testing
xuantengh
committed
0dd7928
add head perm
xuantengh
committed
5ad0a44
fix test
xuantengh
committed
a188b24
move unused function to test file
xuantengh
committed
aa3fbf9
update comments
xuantengh
committed
757612e
move head_perm to forward
xuantengh
committed
2a80dcf
add lru cache
xuantengh
committed
a2bf804
type annotation
xuantengh
committed
41a848f
fix tuple
xuantengh
committed
859a6dc
fix test nvls
xuantengh
committed
5d261d1
Merge branch 'main' into xuantengh/gdn_a2a
xuantengh
authored
4137f93