Support multi-node DP attention (#2925)

Co-authored-by: dhou-xai <dhou@x.ai>

python/sglang/srt/models/deepseek_v2.py

@@ -855,10 +855,9 @@ class DeepseekV2ForCausalLM(nn.Module):
         forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, forward_batch)
-        if not forward_batch.forward_mode.is_idle():
-            return self.logits_processor(
-                input_ids, hidden_states, self.lm_head, forward_batch
-            )
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [