NPU Adaption for Sanna

Author: J石页

src/diffusers/models/attention_processor.py

@@ -3159,7 +3159,7 @@ def __call__(
             # scaled_dot_product_attention expects attention_mask shape to be
             # (batch, heads, source_length, target_length)
             attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
-            attention_mask = attention_mask.repeat_interleave(hidden_states.shape[1], dim=2)
+            attention_mask = attention_mask.repeat(1, 1, hidden_states.shape[1], 1)
             if attention_mask.dtype == torch.bool:
                 attention_mask = torch.logical_not(attention_mask.bool())
             else: