[CI fix] Disable paged attn with PT Flash (fairinternal/xformers#1174)

Co-authored-by: danthe3rd <danthe3rd> __original_commit__ = fairinternal/xformers@d4a5d82
facebookresearch · Aug 6, 2024 · 567cb66 · 567cb66
1 parent fd6821a
commit 567cb66
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/xformers/ops/fmha/flash.py b/xformers/ops/fmha/flash.py
@@ -489,7 +489,7 @@ def _is_causal(attn_bias: Optional[Union[torch.Tensor, AttentionBias]]) -> bool:
 
 def _is_paged_attention_supported(attn_bias_type) -> bool:
     if issubclass(attn_bias_type, PagedBlockDiagonalPaddedKeysMask):
-        return FLASH_VERSION > "2.5.6"
+        return FLASH_VERSION > "2.5.6" and not _USE_PT_FLASH_ATTN
 
     return True