A high-throughput and memory-efficient inference and serving engine for LLMs
[Bugfix][Model] Qwen3-Omni: move cu_seqlens to GPU before VIT attention (#44264)
Signed-off-by: Lanze Liu <lanzetech@gmail.com>
Lanze Liu committed
Jun 8, 2026 at 20:02 UTC
540aaf21406bad1eb0f3d5ddc5e835e45ea01e43
1 file changed
+3
vllm/model_executor/models/qwen3_omni_moe_thinker.py
+3
@@ -991,6 +991,9 @@ class Qwen3Omni_VisionTransformer(nn.Module):
991
)
992
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
993
994
+ # Move cu_seqlens to GPU; grid_thw may be on CPU during profile_run
995
+ # and FA3 vit attention requires cu_seqlens on CUDA.
996
+ cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
997
hidden_states = hidden_states.unsqueeze(1)
998
rotary_pos_emb_cos = rotary_pos_emb_cos.to(hidden_states.device)
999
rotary_pos_emb_sin = rotary_pos_emb_sin.to(hidden_states.device)