backwards pass is not recommended under the autocast context, per pytorch docs

dalle2_pytorch/train.py

@@ -264,8 +264,8 @@ class DiffusionPriorTrainer(nn.Module):
         for chunk_size_frac, (chunked_args, chunked_kwargs) in split_args_and_kwargs(x, *args, split_size = max_batch_size, **kwargs):
             with autocast(enabled = self.amp):
                 loss = self.diffusion_prior(*chunked_args, **chunked_kwargs)
-                loss = loss * chunk_size_frac
 
+            loss = loss * chunk_size_frac
             total_loss += loss.item()
             self.scaler.scale(loss).backward()
 
@@ -388,8 +388,8 @@ class DecoderTrainer(nn.Module):
         for chunk_size_frac, (chunked_args, chunked_kwargs) in split_args_and_kwargs(x, split_size = max_batch_size, **kwargs):
             with autocast(enabled = self.amp):
                 loss = self.decoder(*chunked_args, unet_number = unet_number, **chunked_kwargs)
-                loss = loss * chunk_size_frac
 
+            loss = loss * chunk_size_frac
             total_loss += loss.item()
             self.scale(loss, unet_number = unet_number).backward()
 

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.2.29',
+  version = '0.2.28',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',