Migrate logits computation and gather to model_runner (#3233)

vllm/model_executor/models/phi.py

@@ -49,6 +49,7 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                QKVParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.rotary_embedding import get_rope
+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)
@@ -240,7 +241,8 @@ class PhiForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size,
                                       config.hidden_size,
                                       bias=True)
-        self.sampler = Sampler(config.vocab_size)
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = Sampler()
 
     def forward(
         self,
@@ -254,14 +256,18 @@ class PhiForCausalLM(nn.Module):
 
         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, self.lm_head.bias)
+        return logits
+
     def sample(
         self,
-        hidden_states: torch.Tensor,
+        logits: torch.Tensor,
         sampling_metadata: SamplingMetadata,
     ) -> Optional[SamplerOutput]:
-        head = self.lm_head
-        next_tokens = self.sampler(head.weight, hidden_states,
-                                   sampling_metadata, head.bias)
+        next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
     def load_weights(self,