Migrate logits computation and gather to model_runner (#3233)

vllm/model_executor/models/gpt_bigcode.py

@@ -31,6 +31,7 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                LinearMethodBase,
                                                QKVParallelLinear,
                                                RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.sampler import Sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
@@ -237,7 +238,8 @@ class GPTBigCodeForCausalLM(nn.Module):
         self.linear_method = linear_method
         self.transformer = GPTBigCodeModel(config, linear_method)
         self.lm_head_weight = self.transformer.wte.weight
-        self.sampler = Sampler(config.vocab_size)
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = Sampler()
 
     def forward(
         self,
@@ -250,13 +252,18 @@ class GPTBigCodeForCausalLM(nn.Module):
                                          input_metadata)
         return hidden_states
 
+    def compute_logits(self, hidden_states: torch.Tensor,
+                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
+        logits = self.logits_processor(self.lm_head_weight, hidden_states,
+                                       sampling_metadata)
+        return logits
+
     def sample(
         self,
-        hidden_states: torch.Tensor,
+        logits: torch.Tensor,
         sampling_metadata: SamplingMetadata,
     ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(self.lm_head_weight, hidden_states,
-                                   sampling_metadata)
+        next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
     def load_weights(self,