Use NCCL instead of ray for control-plane communication to remove serialization overhead (#2221)

vllm/model_executor/models/gpt_bigcode.py

@@ -254,7 +254,7 @@ class GPTBigCodeForCausalLM(nn.Module):
         self,
         hidden_states: torch.Tensor,
         sampling_metadata: SamplingMetadata,
-    ) -> SamplerOutput:
+    ) -> Optional[SamplerOutput]:
         next_tokens = self.sampler(self.lm_head_weight, hidden_states,
                                    sampling_metadata)
         return next_tokens