EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

bugfix: Fix output_ids extraction in detokenizer_manager (#9047)

Browse Source
This commit is contained in:
Chang Su
2025-08-11 03:17:32 -07:00
committed by GitHub
parent f4ae50e97c
commit a6452b7188
2 changed files with 2 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
python/sglang/srt/entrypoints/context.py
View File

@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0 # 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 json
import logging import logging
from abc import ABC, abstractmethod from abc import ABC, abstractmethod
@@ -83,14 +83,6 @@ class HarmonyContext(ConversationContext):
if isinstance(output, dict) and "output_ids" in output: if isinstance(output, dict) and "output_ids" in output:
output_token_ids = output["output_ids"] 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: for token_id in output_token_ids:
self.parser.process(token_id) self.parser.process(token_id)
output_msgs = self.parser.messages output_msgs = self.parser.messages
@@ -196,15 +188,6 @@ class StreamingHarmonyContext(HarmonyContext):
# RequestOutput from SGLang with outputs # RequestOutput from SGLang with outputs
output_token_ids = output["output_ids"] 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: for token_id in output_token_ids:
self.parser.process(token_id) self.parser.process(token_id)

2
python/sglang/srt/managers/detokenizer_manager.py
View File

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