onnxruntime
47716d36 - FlashAttention: use subgroup size 32 on Intel Xe2 and Xe3 (#33808)

Commit
1 day ago
FlashAttention: use subgroup size 32 on Intel Xe2 and Xe3 (#33808) An explicit subgroup size lets Tint use a constant lane clamp bound for subgroupShuffle, so IGC keeps direct register addressing.
Author
Parents
Loading