add fa unified attn 导入判断

vllm/v1/attention/ops/triton_unified_attention.py

@@ -13,9 +13,10 @@ from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
 from vllm import envs
-from flash_attn import (
-    varlen_fwd_unified,
-)
+try:
+    from flash_attn import varlen_fwd_unified
+except Exception:
+    varlen_fwd_unified = None
 
 logger = init_logger(__name__)
 float8_info = torch.finfo(current_platform.fp8_dtype())
@@ -1045,6 +1046,10 @@ def unified_attention(
                 USE_FP8=output_scale is not None,
             )
         else:
+            if varlen_fwd_unified is None:
+                raise RuntimeError(
+                    "flash_attn.varlen_fwd_unified is not available in this flash-attn version"
+                )
             # print("Running FA kernel")
             varlen_fwd_unified(
                 q=q,