[MM Encoder] Add Triton ViT attention backend (#32183)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-02-15 06:32:47 -08:00
committed by GitHub
parent 19fab44152
commit 71cd89264f
14 changed files with 178 additions and 51 deletions
+5 -4
View File
@@ -723,10 +723,11 @@ class Glm4vVisionTransformer(nn.Module):
cu_seqlens: torch.Tensor,
) -> torch.Tensor | 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