NPU Adaption for Sanna

Author: J石页

src/diffusers/models/attention_processor.py

@@ -3158,10 +3158,12 @@ def __call__(
             attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
             # scaled_dot_product_attention expects attention_mask shape to be
             # (batch, heads, source_length, target_length)
-            attn_mask = attention_mask[0]
-            seq_len = hidden_states.shape[1]
-            attention_mask = attn_mask.repeat_interleave(seq_len * batch_size, dim=0)
             attention_mask = attention_mask.view(batch_size, 1, -1, attention_mask.shape[-1])
+            attention_mask = attention_mask.repeat_interleave(hidden_states.shape[1], dim=2)
+            if attention_mask.dtype == torch.bool:
+                attention_mask = torch.logical_not(attention_mask.bool())
+            else:
+                attention_mask = attention_mask.bool()
 
             if attention_mask.dtype != torch.uint8:
                 if attention_mask.dtype == torch.bool: