[MM Encoder] Add Triton ViT attention backend (#32183)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
@@ -573,10 +573,11 @@ class DotsVisionTransformer(nn.Module):
|
||||
|
||||
def compute_attn_mask_seqlen(self, cu_seqlens: torch.Tensor) -> int | None:
|
||||
max_seqlen = None
|
||||
if (
|
||||
self.attn_backend == AttentionBackendEnum.FLASH_ATTN
|
||||
or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
|
||||
):
|
||||
if self.attn_backend in {
|
||||
AttentionBackendEnum.FLASH_ATTN,
|
||||
AttentionBackendEnum.ROCM_AITER_FA,
|
||||
AttentionBackendEnum.TRITON_ATTN,
|
||||
}:
|
||||
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
|
||||
return max_seqlen
|
||||
|
||||
|
||||
Reference in New Issue
Block a user