merge v0.3.1

vllm/entrypoints/llm.py

@@ -111,13 +111,13 @@ class LLM:
 
     def get_tokenizer(
             self) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
-        return self.llm_engine.tokenizer
+        return self.llm_engine.tokenizer.tokenizer
 
     def set_tokenizer(
         self,
         tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
     ) -> None:
-        self.llm_engine.tokenizer = tokenizer
+        self.llm_engine.tokenizer.tokenizer = tokenizer
 
     def generate(
         self,

vllm/model_executor/layers/activation.py

@@ -89,9 +89,7 @@ class ScaledActivation(nn.Module):
         if params_dtype is None:
             params_dtype = torch.get_default_dtype()
         self.scales = nn.Parameter(
-            torch.empty(intermediate_size_per_partition,
-                        dtype=params_dtype,
-                        device="cuda"))
+            torch.empty(intermediate_size_per_partition, dtype=params_dtype))
         set_weight_attrs(self.scales, {"weight_loader": self.weight_loader})
 
     def forward(self, x: torch.Tensor) -> torch.Tensor: