Fix FA4 dynamic_causal for full attention layers (#46659)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Jun 29, 2026 at 17:23 UTC
8fc1b2d046f4a991b5c24bb5470bf45efcfe9d01
1 file changed
+4
-1
vllm/v1/attention/backends/flash_attn.py
+4
-1
@@ -906,7 +906,10 @@ class FlashAttentionImpl(AttentionImpl):
906
f"FA{self.vllm_flash_attn_version}"
907
)
908
dynamic_causal = causal
909
- causal = False
909
+ has_window = (
910
+ sliding_window_size is not None and sliding_window_size[1] >= 0
911
+ )
912
+ causal = not has_window
913
914
flash_attn_varlen_func(
915
q=query[:num_actual_tokens],