Upload folder using huggingface_hub

animatediff/models/motion_module.py

@@ -308,9 +308,14 @@ class VersatileAttention(CrossAttention):
                 attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
                 attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)
 
+        if not hasattr(self, '_use_memory_efficient_attention_xformers'):
+            self._use_memory_efficient_attention_xformers = True
+
+
         # attention, what we cannot get enough of
         if self._use_memory_efficient_attention_xformers:
-            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+            self.set_use_memory_efficient_attention_xformers(True)
+            # hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
             # Some versions of xformers return output in fp32, cast it back to the dtype of the input
             hidden_states = hidden_states.to(query.dtype)
         else:

train.py

@@ -177,6 +177,7 @@ def main(
     for name, module in unet.named_modules():
         if "motion_modules" in name and (train_whole_module or name.endswith(tuple(trainable_modules))):
             for params in module.parameters():
+                print("trainable", name)
                 params.requires_grad = True
 
     if enable_xformers_memory_efficient_attention:
@@ -370,10 +371,21 @@ def main(
                 avg_loss = accelerator.gather(loss.repeat(train_batch_size)).mean()
                 train_loss += avg_loss.item() / gradient_accumulation_steps
 
+                for name, module in unet.named_modules():
+                    if "motion_modules" in name and (train_whole_module or name.endswith(tuple(trainable_modules))):
+                        for params in module.parameters():
+                            params.requires_grad = True
+
                 # Backpropagate
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
                     accelerator.clip_grad_norm_(unet.parameters(), max_grad_norm)
+                    
+                # for param in unet.parameters():
+                    # print(param.grad)
+                    
+                    
+                    
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()