Refactor Worker & InputMetadata (#1843)

vllm/model_executor/models/gpt_bigcode.py

@@ -36,6 +36,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.parallel_utils.parallel_state import (
     get_tensor_model_parallel_world_size)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.weight_utils import (default_weight_loader,
                                               hf_model_weights_iterator)
 from vllm.sequence import SamplerOutput
@@ -251,11 +252,18 @@ class GPTBigCodeForCausalLM(nn.Module):
         kv_caches: List[KVCache],
         input_metadata: InputMetadata,
         cache_events: Optional[List[torch.cuda.Event]],
-    ) -> SamplerOutput:
+    ) -> torch.Tensor:
         hidden_states = self.transformer(input_ids, positions, kv_caches,
                                          input_metadata, cache_events)
+        return hidden_states
+
+    def sample(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> SamplerOutput:
         next_tokens = self.sampler(self.lm_head_weight, hidden_states,
-                                   input_metadata)
+                                   sampling_metadata)
         return next_tokens
 
     def load_weights(self,