Add logprobs support with OpenAI API chat completions

docs/openai_api.md

@@ -63,6 +63,24 @@ completion = openai.chat.completions.create(
 print(completion.choices[0].message.content)
 ```
 
+### Logprobs 
+
+Logprobs are supported with the OpenAI API:
+
+```python
+
+# create a chat completion
+completion = openai.chat.completions.create(
+  model=model,
+  messages=[{"role": "user", "content": "Hello! What is your name?"}],
+  log_probs=True
+)
+# print the completion
+print(completion.choices[0].logprobs)
+```
+
+### Streaming
+
 Streaming is also supported. See [test_openai_api.py](../tests/test_openai_api.py).  If your api server is behind a proxy you'll need to turn off buffering, you can do so in Nginx by setting `proxy_buffering off;` in the location block for the proxy.
 
 ### cURL
@@ -146,7 +164,7 @@ export FASTCHAT_WORKER_API_EMBEDDING_BATCH_SIZE=1
 ## Todos
 Some features to be implemented:
 
-- [ ] Support more parameters like `logprobs`, `logit_bias`, `user`, `presence_penalty` and `frequency_penalty`
+- [ ] Support more parameters like `logit_bias`, `user`, `presence_penalty` and `frequency_penalty`
 - [ ] Model details (permissions, owner and create time)
 - [ ] Edits API
 - [ ] Rate Limitation Settings

fastchat/protocol/openai_api_protocol.py

@@ -64,6 +64,7 @@ class ChatCompletionRequest(BaseModel):
     ]
     temperature: Optional[float] = 0.7
     top_p: Optional[float] = 1.0
+    logprobs: Optional[int] = None
     top_k: Optional[int] = -1
     n: Optional[int] = 1
     max_tokens: Optional[int] = None
@@ -81,6 +82,7 @@ class ChatMessage(BaseModel):
 
 class ChatCompletionResponseChoice(BaseModel):
     index: int
+    logprobs: Optional[LogProbs] = None
     message: ChatMessage
     finish_reason: Optional[Literal["stop", "length"]] = None
 

fastchat/serve/openai_api_server.py

@@ -275,7 +275,7 @@ async def get_gen_params(
     frequency_penalty: Optional[float],
     max_tokens: Optional[int],
     echo: Optional[bool],
-    logprobs: Optional[int] = None,
+    logprobs: Optional[int],
     stop: Optional[Union[str, List[str]]],
     best_of: Optional[int] = None,
     use_beam_search: Optional[bool] = None,
@@ -431,6 +431,7 @@ async def create_chat_completion(request: ChatCompletionRequest):
         frequency_penalty=request.frequency_penalty,
         max_tokens=request.max_tokens,
         echo=False,
+        logprobs=request.logprobs,
         stop=request.stop,
     )
 
@@ -472,6 +473,7 @@ async def create_chat_completion(request: ChatCompletionRequest):
             ChatCompletionResponseChoice(
                 index=i,
                 message=ChatMessage(role="assistant", content=content["text"]),
+                logprobs=create_openai_logprobs(content.get("logprobs", None)),
                 finish_reason=content.get("finish_reason", "stop"),
             )
         )