Migrate logits computation and gather to model_runner (#3233)

vllm/model_executor/models/starcoder2.py

@@ -32,6 +32,7 @@ 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, ParallelLMHead, DEFAULT_VOCAB_PADDING_SIZE)
@@ -254,7 +255,9 @@ class Starcoder2ForCausalLM(nn.Module):
                 padding_size=DEFAULT_VOCAB_PADDING_SIZE,
             )
             self.lm_head_weight = self.lm_head.weight
-        self.sampler = Sampler(self.unpadded_vocab_size, config.vocab_size)
+        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                config.vocab_size)
+        self.sampler = Sampler()
 
     def forward(
         self,
@@ -267,13 +270,18 @@ class Starcoder2ForCausalLM(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: Optional[torch.Tensor],
+        logits: Optional[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,