be more aggressive at clearing out saved_attn_slice

ldm/models/diffusion/cross_attention_control.py

@@ -116,6 +116,7 @@ class CrossAttentionControl:
         for m in self_attention_modules+tokens_attention_modules:
             m.save_last_attn_slice = False
             m.use_last_attn_slice = False
+            m.last_attn_slice = None
 
     @classmethod
     def request_save_attention_maps(cls, model, cross_attention_type: CrossAttentionType):

ldm/models/diffusion/shared_invokeai_diffusion.py

@@ -134,6 +134,8 @@ class InvokeAIDiffuserComponent:
         # representing batched uncond + cond, but then when it comes to applying the saved attention, the
         # wrangler gets an attention tensor which only has shape[0]=8, representing just self.edited_conditionings.)
         # todo: give CrossAttentionControl's `wrangler` function more info so it can work with a batched call as well.
+
+        try:
             unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning)
 
             # process x using the original prompt, saving the attention maps
@@ -152,6 +154,12 @@ class InvokeAIDiffuserComponent:
 
             return unconditioned_next_x, conditioned_next_x
 
+        except RuntimeError:
+            # make sure we clean out the attention slices we're storing on the model
+            # TODO don't store things on the model
+            CrossAttentionControl.clear_requests(self.model)
+            raise
+
     def estimate_percent_through(self, step_index, sigma):
         if step_index is not None and self.cross_attention_control_context is not None:
             # percent_through will never reach 1.0 (but this is intended)