[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
+2 -8
View File
@@ -730,14 +730,7 @@ class SiglipEncoder(nn.Module):
head_size=head_dim,
dtype=torch.get_default_dtype(),
)
if self.attn_backend not in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.ROCM_AITER_FA,
}:
raise RuntimeError(
f"PaddleOCR-VL does not support {self.attn_backend} backend now."
)
self.layers = nn.ModuleList(
[
SiglipEncoderLayer(
@@ -805,6 +798,7 @@ class SiglipEncoder(nn.Module):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()