bugfix: Fix output_ids extraction in detokenizer_manager (#9047)

python/sglang/srt/entrypoints/context.py

 # SPDX-License-Identifier: Apache-2.0
-# Copied from vLLM
+# Copied from vLLM: https://github.com/zyongye/vllm/blob/6a70830065701b163e36a86fd331b41b5feac401/vllm/entrypoints/context.py
 import json
 import logging
 from abc import ABC, abstractmethod
@@ -83,14 +83,6 @@ class HarmonyContext(ConversationContext):
         if isinstance(output, dict) and "output_ids" in output:
             output_token_ids = output["output_ids"]
 
-            # TODO: REMOVE here:
-            # Very hacky, find the first occurrence of token 200006 and cut from there
-            try:
-                start_index = output_token_ids.index(200006)
-                output_token_ids = output_token_ids[start_index:]
-            except ValueError:
-                pass
-
             for token_id in output_token_ids:
                 self.parser.process(token_id)
             output_msgs = self.parser.messages
@@ -196,15 +188,6 @@ class StreamingHarmonyContext(HarmonyContext):
             # RequestOutput from SGLang with outputs
             output_token_ids = output["output_ids"]
 
-            # TODO: REMOVE here:
-            # Very hacky, find the first occurrence of token 200006 and cut from there
-            # Find the first occurrence of token 200006 and cut from there
-            try:
-                start_index = output_token_ids.index(200006)
-                output_token_ids = output_token_ids[start_index:]
-            except ValueError:
-                pass
-
             for token_id in output_token_ids:
                 self.parser.process(token_id)
 

python/sglang/srt/managers/detokenizer_manager.py

@@ -216,7 +216,7 @@ class DetokenizerManager:
             rids=recv_obj.rids,
             finished_reasons=recv_obj.finished_reasons,
             output_strs=output_strs,
-            output_ids=recv_obj.decode_ids,
+            output_ids=recv_obj.output_ids,
             prompt_tokens=recv_obj.prompt_tokens,
             completion_tokens=recv_obj.completion_tokens,
             cached_tokens=recv_obj.cached_tokens,