FIX [`Generation`] Fix some issues when running the MaxLength criteria on CPU (#29317)

fix the bitwise or issue

src/transformers/generation/stopping_criteria.py

@@ -73,7 +73,7 @@ class MaxLengthCriteria(StoppingCriteria):
                 f"maximum length ({self.max_position_embeddings}). Depending on the model, you may observe "
                 "exceptions, performance degradation, or nothing at all."
             )
-        return torch.full((input_ids.shape[0],), is_done, device=input_ids.device)
+        return torch.full((input_ids.shape[0],), is_done, device=input_ids.device, dtype=torch.bool)
 
 
 class MaxNewTokensCriteria(StoppingCriteria):
@@ -103,7 +103,7 @@ class MaxNewTokensCriteria(StoppingCriteria):
     @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> torch.BoolTensor:
         is_done = input_ids.shape[-1] >= self.max_length
-        return torch.full((input_ids.shape[0],), is_done, device=input_ids.device)
+        return torch.full((input_ids.shape[0],), is_done, device=input_ids.device, dtype=torch.bool)
 
 
 class MaxTimeCriteria(StoppingCriteria):
@@ -126,7 +126,7 @@ class MaxTimeCriteria(StoppingCriteria):
     @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> torch.BoolTensor:
         is_done = time.time() - self.initial_timestamp > self.max_time
-        return torch.full((input_ids.shape[0],), is_done, device=input_ids.device)
+        return torch.full((input_ids.shape[0],), is_done, device=input_ids.device, dtype=torch.bool)
 
 
 class StoppingCriteriaList(list):