Update distributed.py (bigscience-workshop#275)

Author: zhangsheng377

megatron/model/distributed.py

@@ -217,7 +217,6 @@ def allreduce_gradients(self):
                     if tp not in buckets:
                         buckets[tp] = []
                     buckets[tp].append(param)
-                    param.main_grad = param.grad
 
             # For each bucket, all-reduce and copy all-reduced grads.
             for tp in buckets: