A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 02:15 UTC
[ROCm] [Bugfix] `compute_attn_mask_seqlen` for qwen3 omni (#29974)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
TJian committed
Dec 4, 2025 at 16:20 UTC
3f1b03739ae1422361446d3d23bed970bd549ebc
1 file changed
+4
-1
vllm/model_executor/models/qwen3_omni_moe_thinker.py
+4
-1
@@ -494,7 +494,10 @@ class Qwen3Omni_VisionTransformer(nn.Module):
494
cu_seqlens: torch.Tensor,
495
) -> torch.Tensor:
496
max_seqlen = torch.zeros([], device=cu_seqlens.device)
497
- if self.attn_backend == AttentionBackendEnum.FLASH_ATTN:
497
+ if self.attn_backend in {
498
+ AttentionBackendEnum.FLASH_ATTN,
499
+ AttentionBackendEnum.ROCM_AITER_FA,
500
+ }:
501
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
502
return max_seqlen
503