NPU Adaption for Sanna

Author: J石页

src/diffusers/models/attention_processor.py

@@ -294,10 +294,6 @@ def __init__(
             processor = (
                 AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
             )
-
-        if is_torch_npu_available():
-            if isinstance(processor, AttnProcessor2_0):
-                processor = AttnProcessorNPU()
         self.set_processor(processor)
 
     def set_use_xla_flash_attention(
@@ -525,6 +521,11 @@ def set_processor(self, processor: "AttnProcessor") -> None:
             processor (`AttnProcessor`):
                 The attention processor to use.
         """
+        # set to use npu flash attention from 'torch_npu' if available
+        if is_torch_npu_available():
+            if isinstance(processor, AttnProcessor2_0):
+                processor = AttnProcessorNPU()
+
         # if current processor is in `self._modules` and if passed `processor` is not, we need to
         # pop `processor` from `self._modules`
         if (