fix grad norm bug

src/twinkle/infra/__init__.py

@@ -5,7 +5,9 @@
 import os
 from typing import Any, Callable, List, Literal, Optional, TypeVar, Union
 
-from twinkle.utils import DeviceGroup, DeviceMesh, Platform, check_unsafe, framework_util, requires
+from twinkle.utils import DeviceGroup, DeviceMesh, Platform, check_unsafe, framework_util, get_logger, requires
+
+logger = get_logger()
 
 T1 = TypeVar('T1', bound=object)
 
@@ -437,6 +439,9 @@ def new_init(self, *args, **kwargs):
                 # Pass an instance_id is recommended
                 instance_id = kwargs.pop('instance_id', '') + f'{caller_file}_{caller_line}'
                 remote_group = kwargs.get('remote_group')
+                if remote_group is None:
+                    logger.info(f'⚠️ Using local initialization of class: {cls}, please make sure the class '
+                                'does not need remote execution.')
                 # If cannot trust_remote_code, no callable and type can be used.
                 check_unsafe(*args, **kwargs)
 

src/twinkle/loss/cross_entropy.py

@@ -11,4 +11,8 @@ def __call__(self, inputs, outputs, **kwargs):
         import torch
         logits = outputs['logits'].view(-1, outputs['logits'].shape[-1])
         labels = inputs['labels'].view(-1)
-        return torch.nn.CrossEntropyLoss(reduction=self.reduction)(logits, labels)
+        loss = torch.nn.CrossEntropyLoss(reduction=self.reduction)(logits, labels)
+        if self.reduction != 'sum':
+            return loss
+        else:
+            return loss, (labels != -100).sum()