Use reduce scatter for DP (#8539)

python/sglang/srt/layers/linear.py

@@ -1277,7 +1277,7 @@ class RowParallelLinear(LinearBase):
             # It does not support additional parameters.
             param.load_row_parallel_weight(loaded_weight)
 
-    def forward(self, input_, can_fuse_mlp_allreduce=False):
+    def forward(self, input_, skip_all_reduce=False):
         if self.input_is_parallel:
             input_parallel = input_
         else:
@@ -1294,7 +1294,7 @@ class RowParallelLinear(LinearBase):
         with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
             output_parallel = self.quant_method.apply(self, input_parallel, bias=bias_)
             sm.tag(output_parallel)
-        if self.reduce_results and self.tp_size > 1 and not can_fuse_mlp_allreduce:
+        if self.reduce_results and self.tp_size > 1 and not skip_all_reduce:
             output = tensor_model_parallel_all_reduce(output_parallel)
         else:
             output = output_parallel