biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-02-15 22:32:47 +08:00
committed by GitHub
parent 19fab44152
commit 71cd89264f
14 changed files with 178 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/models/qwen3_omni_moe_thinker.py
View File

@@ -391,6 +391,7 @@ class Qwen3OmniMoeAudioEncoder(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()
return max_seqlen
@@ -919,6 +920,7 @@ class Qwen3Omni_VisionTransformer(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()
return max_seqlen