Add support for logprobs in OpenAI chat API (#852)

examples/usage/openai_parallel_sample.py

@@ -106,13 +106,12 @@ response = client.chat.completions.create(
         {"role": "user", "content": "List 3 countries and their capitals."},
     ],
     temperature=0.8,
-    max_tokens=64,
+    max_tokens=1,
     logprobs=True,
-    n=1,
+    top_logprobs=3,
 )
 print(response)
 
-
 # Chat completion
 response = client.chat.completions.create(
     model="default",
@@ -121,8 +120,34 @@ response = client.chat.completions.create(
         {"role": "user", "content": "List 3 countries and their capitals."},
     ],
     temperature=0.8,
-    max_tokens=64,
+    max_tokens=1,
+    n=1,
+)
+print(response)
+
+# Chat completion
+response = client.chat.completions.create(
+    model="default",
+    messages=[
+        {"role": "system", "content": "You are a helpful AI assistant"},
+        {"role": "user", "content": "List 3 countries and their capitals."},
+    ],
+    temperature=0.8,
+    max_tokens=1,
     logprobs=True,
+    top_logprobs=3,
+)
+print(response)
+
+# Chat completion
+response = client.chat.completions.create(
+    model="default",
+    messages=[
+        {"role": "system", "content": "You are a helpful AI assistant"},
+        {"role": "user", "content": "List 3 countries and their capitals."},
+    ],
+    temperature=0.8,
+    max_tokens=1,
     n=4,
 )
 print(response)

python/sglang/srt/openai_api/adapter.py

@@ -43,7 +43,9 @@ from sglang.srt.openai_api.protocol import (
     ChatCompletionResponseChoice,
     ChatCompletionResponseStreamChoice,
     ChatCompletionStreamResponse,
+    ChatCompletionTokenLogprob,
     ChatMessage,
+    ChoiceLogprobs,
     CompletionRequest,
     CompletionResponse,
     CompletionResponseChoice,
@@ -54,6 +56,7 @@ from sglang.srt.openai_api.protocol import (
     FileRequest,
     FileResponse,
     LogProbs,
+    TopLogprob,
     UsageInfo,
 )
 
@@ -70,7 +73,7 @@ class FileMetadata:
 batch_storage: Dict[str, BatchResponse] = {}
 file_id_request: Dict[str, FileMetadata] = {}
 file_id_response: Dict[str, FileResponse] = {}
-## map file id to file path in SGlang backend
+# map file id to file path in SGlang backend
 file_id_storage: Dict[str, str] = {}
 
 
@@ -261,7 +264,7 @@ async def process_batch(tokenizer_manager, batch_id: str, batch_request: BatchRe
             failed_requests += len(file_request_list)
 
         for idx, response in enumerate(responses):
-            ## the batch_req here can be changed to be named within a batch granularity
+            # the batch_req here can be changed to be named within a batch granularity
             response_json = {
                 "id": f"batch_req_{uuid.uuid4()}",
                 "custom_id": file_request_list[idx].get("custom_id"),
@@ -333,6 +336,8 @@ def v1_generate_request(all_requests):
 
     prompts = []
     sampling_params_list = []
+    return_logprobs = []
+    top_logprobs_nums = []
     first_prompt_type = type(all_requests[0].prompt)
     for request in all_requests:
         prompt = request.prompt
@@ -340,6 +345,10 @@ def v1_generate_request(all_requests):
             type(prompt) == first_prompt_type
         ), "All prompts must be of the same type in file input settings"
         prompts.append(prompt)
+        return_logprobs.append(request.logprobs is not None and request.logprobs > 0)
+        top_logprobs_nums.append(
+            request.logprobs if request.logprobs is not None else 0
+        )
         sampling_params_list.append(
             {
                 "temperature": request.temperature,
@@ -361,6 +370,8 @@ def v1_generate_request(all_requests):
     if len(all_requests) == 1:
         prompt = prompts[0]
         sampling_params_list = sampling_params_list[0]
+        return_logprobs = return_logprobs[0]
+        top_logprobs_nums = top_logprobs_nums[0]
         if isinstance(prompt, str) or isinstance(prompt[0], str):
             prompt_kwargs = {"text": prompt}
         else:
@@ -370,15 +381,11 @@ def v1_generate_request(all_requests):
             prompt_kwargs = {"text": prompts}
         else:
             prompt_kwargs = {"input_ids": prompts}
-
     adapted_request = GenerateReqInput(
         **prompt_kwargs,
         sampling_params=sampling_params_list,
-        return_logprob=all_requests[0].logprobs is not None
+        return_logprob=return_logprobs,
-        and all_requests[0].logprobs > 0,
+        top_logprobs_num=top_logprobs_nums,
-        top_logprobs_num=(
-            all_requests[0].logprobs if all_requests[0].logprobs is not None else 0
-        ),
         return_text_in_logprobs=True,
         stream=all_requests[0].stream,
     )
@@ -430,7 +437,7 @@ def v1_generate_response(request, ret, to_file=False):
             logprobs = None
 
         if to_file:
-            ## to make the choise data json serializable
+            # to make the choise data json serializable
             choice_data = {
                 "index": 0,
                 "text": text,
@@ -454,7 +461,7 @@ def v1_generate_response(request, ret, to_file=False):
                 "status_code": 200,
                 "request_id": ret[i]["meta_info"]["id"],
                 "body": {
-                    ## remain the same but if needed we can change that
+                    # remain the same but if needed we can change that
                     "id": ret[i]["meta_info"]["id"],
                     "object": "text_completion",
                     "created": int(time.time()),
@@ -590,6 +597,8 @@ def v1_chat_generate_request(all_requests, tokenizer_manager):
     texts = []
     sampling_params_list = []
     image_data_list = []
+    return_logprobs = []
+    top_logprobs_nums = []
     for request in all_requests:
         # Prep the data needed for the underlying GenerateReqInput:
         #  - prompt: The full prompt string.
@@ -620,6 +629,8 @@ def v1_chat_generate_request(all_requests, tokenizer_manager):
             stop = request.stop
             image_data = None
         texts.append(prompt)
+        return_logprobs.append(request.logprobs)
+        top_logprobs_nums.append(request.top_logprobs)
         sampling_params_list.append(
             {
                 "temperature": request.temperature,
@@ -637,11 +648,16 @@ def v1_chat_generate_request(all_requests, tokenizer_manager):
         texts = texts[0]
         sampling_params_list = sampling_params_list[0]
         image_data = image_data_list[0]
+        return_logprobs = return_logprobs[0]
+        top_logprobs_nums = top_logprobs_nums[0]
     adapted_request = GenerateReqInput(
         text=texts,
         image_data=image_data,
         sampling_params=sampling_params_list,
-        stream=request.stream,
+        return_logprob=return_logprobs,
+        top_logprobs_num=top_logprobs_nums,
+        stream=all_requests[0].stream,
+        return_text_in_logprobs=True,
     )
     if len(all_requests) == 1:
         return adapted_request, all_requests[0]
@@ -654,26 +670,63 @@ def v1_chat_generate_response(request, ret, to_file=False):
     total_completion_tokens = 0
 
     for idx, ret_item in enumerate(ret):
+        logprobs = False
+        if isinstance(request, List) and request[idx].logprobs:
+            logprobs = True
+        elif (not isinstance(request, List)) and request.logprobs:
+            logprobs = True
+        if logprobs:
+            logprobs = to_openai_style_logprobs(
+                output_token_logprobs=ret_item["meta_info"]["output_token_logprobs"],
+                output_top_logprobs=ret_item["meta_info"]["output_top_logprobs"],
+            )
+            token_logprobs = []
+            for token, logprob in zip(logprobs.tokens, logprobs.token_logprobs):
+                token_bytes = list(token.encode("utf-8"))
+                top_logprobs = []
+                if logprobs.top_logprobs:
+                    for top_token, top_logprob in logprobs.top_logprobs[0].items():
+                        top_token_bytes = list(top_token.encode("utf-8"))
+                        top_logprobs.append(
+                            TopLogprob(
+                                token=top_token,
+                                bytes=top_token_bytes,
+                                logprob=top_logprob,
+                            )
+                        )
+                token_logprobs.append(
+                    ChatCompletionTokenLogprob(
+                        token=token,
+                        bytes=token_bytes,
+                        logprob=logprob,
+                        top_logprobs=top_logprobs,
+                    )
+                )
+
+            choice_logprobs = ChoiceLogprobs(content=token_logprobs)
+        else:
+            choice_logprobs = None
         prompt_tokens = ret_item["meta_info"]["prompt_tokens"]
         completion_tokens = ret_item["meta_info"]["completion_tokens"]
 
         if to_file:
-            ## to make the choice data json serializable
+            # to make the choice data json serializable
             choice_data = {
                 "index": 0,
                 "message": {"role": "assistant", "content": ret_item["text"]},
-                "logprobs": None,
+                "logprobs": choice_logprobs,
                 "finish_reason": ret_item["meta_info"]["finish_reason"],
             }
         else:
             choice_data = ChatCompletionResponseChoice(
                 index=idx,
                 message=ChatMessage(role="assistant", content=ret_item["text"]),
+                logprobs=choice_logprobs,
                 finish_reason=ret_item["meta_info"]["finish_reason"],
             )
 
         choices.append(choice_data)
-        total_prompt_tokens = prompt_tokens
+        total_prompt_tokens += prompt_tokens
         total_completion_tokens += completion_tokens
     if to_file:
         responses = []
@@ -683,7 +736,7 @@ def v1_chat_generate_response(request, ret, to_file=False):
                 "status_code": 200,
                 "request_id": ret[i]["meta_info"]["id"],
                 "body": {
-                    ## remain the same but if needed we can change that
+                    # remain the same but if needed we can change that
                     "id": ret[i]["meta_info"]["id"],
                     "object": "chat.completion",
                     "created": int(time.time()),

python/sglang/srt/openai_api/protocol.py

@@ -54,6 +54,24 @@ class LogProbs(BaseModel):
     top_logprobs: List[Optional[Dict[str, float]]] = Field(default_factory=list)
 
 
+class TopLogprob(BaseModel):
+    token: str
+    bytes: List[int]
+    logprob: float
+
+
+class ChatCompletionTokenLogprob(BaseModel):
+    token: str
+    bytes: List[int]
+    logprob: float
+    top_logprobs: List[TopLogprob]
+
+
+class ChoiceLogprobs(BaseModel):
+    # build for v1/chat/completions response
+    content: List[ChatCompletionTokenLogprob]
+
+
 class UsageInfo(BaseModel):
     prompt_tokens: int = 0
     total_tokens: int = 0
@@ -239,8 +257,8 @@ class ChatMessage(BaseModel):
 class ChatCompletionResponseChoice(BaseModel):
     index: int
     message: ChatMessage
-    logprobs: Optional[LogProbs] = None
+    logprobs: Optional[Union[LogProbs, ChoiceLogprobs]] = None
-    finish_reason: Optional[str] = None
+    finish_reason: str
 
 
 class ChatCompletionResponse(BaseModel):