grad_wei can't be NoneType when running with DeepSpeed, for zero3 wil…

…l divided the gradient

megatron/core/tensor_parallel/layers.py

@@ -285,6 +285,7 @@ def forward(ctx, input, weight, bias, gradient_accumulation_fusion,
     @staticmethod
     @custom_bwd
     def backward(ctx, grad_output):
+        args = get_args()
         input, weight = ctx.saved_tensors
         use_bias = ctx.use_bias
 
@@ -368,7 +369,12 @@ def backward(ctx, grad_output):
         #     grad_weight = grad_output.t().matmul(total_input)
         from megatron.core.tensor_parallel.weight_grad_store import WeightGradStore
         WeightGradStore.put(total_input, grad_output, weight, gradientUpdateFunction)
-        grad_weight = None
+
+        if args.enable_zbh1_pipeline:
+            grad_weight = None
+        else:
+            grad_weight = weight.grad
+
         grad_bias = grad_output.sum(dim=0) if use_bias else None
 
         if ctx.sequence_parallel: