relax_micro_batch_check (#54788)

python/paddle/distributed/fleet/meta_parallel/pipeline_parallel.py

@@ -715,10 +715,7 @@ class PipelineParallel(MetaParallelBase):
             for data in micro_batch_data:
                 self._check_micro_batch_data_valid(data)
         elif micro_batch_data is not None:
-            micro_batch_size = micro_batch_data.shape[0]
-            assert (
-                micro_batch_size == self.micro_batch_size
-            ), f"expected micro_batch_size {self.micro_batch_size} but get {micro_batch_size}"
+            assert isinstance(micro_batch_data, paddle.Tensor)
 
     def _broadcast_final_loss(self):
         # Since the last backward run in interleave will set the virtual rank to 0,