Fix streaming (#600)

python/sglang/backend/runtime_endpoint.py

@@ -180,7 +180,6 @@ class RuntimeEndpoint(BaseBackend):
         self._assert_success(res)
         pos = 0
 
-        incomplete_text = ""
         for chunk in res.iter_lines(decode_unicode=False):
             chunk = chunk.decode("utf-8")
             if chunk and chunk.startswith("data:"):
@@ -188,14 +187,10 @@ class RuntimeEndpoint(BaseBackend):
                     break
                 data = json.loads(chunk[5:].strip("\n"))
                 chunk_text = data["text"][pos:]
-                incomplete_text = data["incomplete_text"]
                 meta_info = data["meta_info"]
                 pos += len(chunk_text)
                 yield chunk_text, meta_info
 
-        if len(incomplete_text) > 0:
-            yield incomplete_text, meta_info
-
     def select(
         self,
         s: StreamExecutor,

python/sglang/srt/managers/detokenizer_manager.py

@@ -55,13 +55,11 @@ class DetokenizerManager:
             # Trim stop str
             # TODO(lmzheng): handle the case where multiple stop strs are hit
             output_strs = []
-            incomplete_strs = []
             for i in range(len(recv_obj.rids)):
                 new_text = read_texts[i][len(surr_texts[i]) :]
-                complete_new_text = find_printable_text(new_text)
-                incomplete_new_text = new_text[len(complete_new_text) :]
-                output_strs.append(recv_obj.decoded_texts[i] + complete_new_text)
-                incomplete_strs.append(incomplete_new_text)
+                if recv_obj.finished_reason[i] is None:
+                    new_text = find_printable_text(new_text)
+                output_strs.append(recv_obj.decoded_texts[i] + new_text)
 
                 if isinstance(recv_obj.finished_reason[i], FINISH_MATCHED_STR):
                     pos = output_strs[i].find(recv_obj.finished_reason[i].matched)
@@ -72,7 +70,6 @@ class DetokenizerManager:
                 BatchStrOut(
                     rids=recv_obj.rids,
                     output_strs=output_strs,
-                    incomplete_strs=incomplete_strs,
                     meta_info=recv_obj.meta_info,
                     finished_reason=recv_obj.finished_reason,
                 )

python/sglang/srt/managers/io_struct.py

@@ -123,7 +123,6 @@ class BatchTokenIDOut:
 class BatchStrOut:
     rids: List[str]
     output_strs: List[str]
-    incomplete_strs: List[str]
     meta_info: List[Dict]
     finished_reason: List[BaseFinishReason]
 

python/sglang/srt/managers/tokenizer_manager.py

@@ -317,7 +317,6 @@ class TokenizerManager:
                 recv_obj.meta_info[i]["id"] = rid
                 out_dict = {
                     "text": recv_obj.output_strs[i],
-                    "incomplete_text": recv_obj.incomplete_strs[i],
                     "meta_info": recv_obj.meta_info[i],
                 }
                 state.out_list.append(out_dict)

python/sglang/srt/openai_api_adapter.py

@@ -164,7 +164,7 @@ async def v1_completions(tokenizer_manager, raw_request: Request):
                         logprobs = None
 
                     delta = text[len(stream_buffer) :]
-                    stream_buffer = content["text"]
+                    stream_buffer = stream_buffer + delta
                     choice_data = CompletionResponseStreamChoice(
                         index=0,
                         text=delta,
@@ -323,7 +323,7 @@ async def v1_chat_completions(tokenizer_manager, raw_request: Request):
 
                     text = content["text"]
                     delta = text[len(stream_buffer) :]
-                    stream_buffer = text
+                    stream_buffer = stream_buffer + delta
                     choice_data = ChatCompletionResponseStreamChoice(
                         index=0,
                         delta=DeltaMessage(content=delta),