[Bugfix] Initialize attention bias on the same device as Query/Key/Value for QwenVL Series (vllm-project#14031)

Author: LouieYang

vllm/model_executor/models/qwen2_5_vl.py

@@ -323,7 +323,8 @@ def forward(
 
             seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
             attn_bias = BlockDiagonalMask.from_seqlens(q_seqlen=seqlens,
-                                                       kv_seqlen=None)
+                                                       kv_seqlen=None,
+                                                       device=q.device)
 
             context_layer = xops.memory_efficient_attention_forward(
                 q, k, v, attn_bias=attn_bias, p=0, scale=None)

vllm/model_executor/models/qwen2_vl.py

@@ -367,7 +367,8 @@ def forward(
 
             seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
             attn_bias = BlockDiagonalMask.from_seqlens(q_seqlen=seqlens,
-                                                       kv_seqlen=None)
+                                                       kv_seqlen=None,
+                                                       device=q.device)
 
             context_layer = xops.memory_efficient_attention_forward(
                 q, k, v, attn_bias=attn_bias, p=0, scale=None)