dont send dummy batch when reloading from checkpoint

also don't crash if param does not recieve grads

fairseq/trainer.py

@@ -271,8 +271,10 @@ class Trainer(object):
             if not p.requires_grad:
                 continue
             if p.grad is None:
-                raise RuntimeError('Model parameter did not receive gradient: ' + name + '. '
-                                   'Use the param in the forward pass or set requires_grad=False')
+                print('WARNING: model parameter did not receive gradient: ' + name + '. '
+                      'Check that you\'re using the param in the forward pass or set requires_grad=False')
+                grads.append(p.new_zeros(p.shape))
+            else:
                 grads.append(p.grad.data)
         return grads
 

train.py

@@ -71,8 +71,7 @@ def main(args):
     )
 
     # Load the latest checkpoint if one is available
-    load_checkpoint(args, trainer, epoch_itr)
-
+    if not load_checkpoint(args, trainer, epoch_itr):
         # Send a dummy batch to warm the caching allocator
         dummy_batch = task.dataset('train').get_dummy_batch(args.max_tokens, max_positions)
         trainer.dummy_train_step(dummy_batch)
@@ -319,6 +318,8 @@ def load_checkpoint(args, trainer, epoch_itr):
             trainer.lr_step_update(trainer.get_num_updates())
             if 'best' in extra_state:
                 save_checkpoint.best = extra_state['best']
+        return True
+    return False
 
 
 def load_dataset_splits(task, splits):