Add the logit_bias option in API serving

src/llamafactory/api/chat.py

@@ -155,6 +155,7 @@ async def create_chat_completion_response(
         max_new_tokens=request.max_tokens,
         num_return_sequences=request.n,
         stop=request.stop,
+        logit_bias=request.logit_bias,
     )
 
     prompt_length, response_length = 0, 0
@@ -214,6 +215,7 @@ async def create_stream_chat_completion_response(
         top_p=request.top_p,
         max_new_tokens=request.max_tokens,
         stop=request.stop,
+        logit_bias=request.logit_bias,
     ):
         if len(new_token) != 0:
             yield _create_stream_chat_completion_chunk(

src/llamafactory/api/protocol.py

@@ -103,6 +103,7 @@ class ChatCompletionRequest(BaseModel):
     max_tokens: Optional[int] = None
     stop: Optional[Union[str, List[str]]] = None
     stream: bool = False
+    logit_bias: Optional[Union[Dict[int, float], Dict[str, float]]] = None
 
 
 class ChatCompletionResponseChoice(BaseModel):

src/llamafactory/chat/hf_engine.py

@@ -117,9 +117,16 @@ def _process_args(
         max_length: Optional[int] = input_kwargs.pop("max_length", None)
         max_new_tokens: Optional[int] = input_kwargs.pop("max_new_tokens", None)
         stop: Optional[Union[str, List[str]]] = input_kwargs.pop("stop", None)
+        logit_bias: Optional[Union[Dict[int, float], Dict[str, float]]] = input_kwargs.pop("logit_bias", None)
 
         if stop is not None:
             logger.warning_rank0("Stop parameter is not supported by the huggingface engine yet.")
+
+        if logit_bias is not None:
+            logit_bias = {
+                int(token): bias
+                for token, bias in logit_bias.items()
+            }
 
         generating_args = generating_args.copy()
         generating_args.update(
@@ -135,6 +142,7 @@ def _process_args(
                 length_penalty=length_penalty if length_penalty is not None else generating_args["length_penalty"],
                 eos_token_id=[tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids,
                 pad_token_id=tokenizer.pad_token_id,
+                logit_bias=logit_bias,
             )
         )
 

src/llamafactory/chat/vllm_engine.py

@@ -140,6 +140,7 @@ async def _generate(
         max_length: Optional[int] = input_kwargs.pop("max_length", None)
         max_new_tokens: Optional[int] = input_kwargs.pop("max_new_tokens", None)
         stop: Optional[Union[str, List[str]]] = input_kwargs.pop("stop", None)
+        logit_bias: Optional[Union[Dict[int, float], Dict[str, float]]] = input_kwargs.pop("logit_bias", None)
 
         if length_penalty is not None:
             logger.warning_rank0("Length penalty is not supported by the vllm engine yet.")
@@ -158,6 +159,12 @@ async def _generate(
         if max_new_tokens:
             max_tokens = max_new_tokens
 
+        if logit_bias:
+            logit_bias = {
+                int(token): bias
+                for token, bias in logit_bias.items()
+            }
+
         sampling_params = SamplingParams(
             n=num_return_sequences,
             repetition_penalty=(
@@ -171,6 +178,7 @@ async def _generate(
             stop_token_ids=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
             max_tokens=max_tokens,
             skip_special_tokens=self.generating_args["skip_special_tokens"],
+            logit_bias=logit_bias,
         )
 
         if images is not None:  # add image features