fix(diffusers): fix bugs of euler discrete scheduler and attention pr…

…ocessor

mindone/diffusers/models/attention_processor.py

@@ -734,7 +734,7 @@ def flash_attention_op(
         # process `attn_mask` as logic is different between PyTorch and Mindspore
         # In MindSpore, False indicates retention and True indicates discard, in PyTorch it is the opposite
         if attn_mask is not None:
-            attn_mask = ops.logical_not(attn_mask) if attn_mask.dtype == ms.bool_ else attn_mask.bool()
+            attn_mask = ops.logical_not(attn_mask)
             attn_mask = ops.broadcast_to(
                 attn_mask, (attn_mask.shape[0], attn_mask.shape[1], query.shape[-2], key.shape[-2])
             )[:, :1, :, :]
@@ -2103,6 +2103,9 @@ def __call__(
         )
 
         if attention_mask is not None:
+            # bool input dtype is not supported in tensor.repeat_interleave
+            if attention_mask.dtype == ms.bool_:
+                attention_mask = attention_mask.to(hidden_states.dtype)
             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)

mindone/diffusers/schedulers/scheduling_euler_discrete.py

@@ -613,7 +613,11 @@ def add_noise(
         # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
         if self.begin_index is None:
             step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+        elif self.step_index is not None:
+            # add_noise is called after first denoising step (for inpainting)
+            step_indices = [self.step_index] * timesteps.shape[0]
         else:
+            # add noise is called before first denoising step to create initial latent(img2img)
             step_indices = [self.begin_index] * timesteps.shape[0]
 
         sigma = sigmas[step_indices].flatten()