[Bugfix] Initialize attention bias on the same device as Query/Key/Value (#13468)

vllm/attention/backends/xformers.py

@@ -673,7 +673,9 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
 
                     # Cross-attention mask is non-causal
                     attn_bias = BlockDiagonalMask.from_seqlens(
-                        attn_metadata.seq_lens, attn_metadata.encoder_seq_lens)
+                        attn_metadata.seq_lens,
+                        attn_metadata.encoder_seq_lens,
+                        device=query.device)
 
                 # Encoder branch of encoder-decoder model uses
                 # attn_metadata.encoder_seq_lens
@@ -683,7 +685,7 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
 
                     # Encoder self-attention mask is non-causal
                     attn_bias = BlockDiagonalMask.from_seqlens(
-                        attn_metadata.encoder_seq_lens)
+                        attn_metadata.encoder_seq_lens, device=query.device)
 
                 # Self-attention block of encoder-only model just
                 # uses the seq_lens directly.
@@ -692,7 +694,7 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
 
                     # Encoder self-attention mask is non-causal
                     attn_bias = BlockDiagonalMask.from_seqlens(
-                        attn_metadata.seq_lens)
+                        attn_metadata.seq_lens, device=query.device)
 
                 # Self-attention block of decoder branch just
                 # uses the seq_lens directly
@@ -701,7 +703,7 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
 
                     # Decoder self-attention mask is causal
                     attn_bias = BlockDiagonalCausalMask.from_seqlens(
-                        attn_metadata.seq_lens)
+                        attn_metadata.seq_lens, device=query.device)
                 else:
                     raise ValueError("Unknown AttentionType: %s", attn_type)