* fix(detokenizer_manager.py): fix truncated decoded output (#586)

Co-authored-by: hnyls2002 <hnyls2002@gmail.com>
This commit is contained in:
Pan Lyu
2024-07-07 05:53:22 +08:00
committed by GitHub
parent c0982ac553
commit 26908d9568
4 changed files with 17 additions and 10 deletions

View File

@@ -11,7 +11,7 @@ from sglang.srt.hf_transformers_utils import get_tokenizer
from sglang.srt.managers.controller.infer_batch import FINISH_MATCHED_STR
from sglang.srt.managers.io_struct import BatchStrOut, BatchTokenIDOut
from sglang.srt.server_args import PortArgs, ServerArgs
from sglang.utils import get_exception_traceback, graceful_registry
from sglang.utils import find_printable_text, get_exception_traceback, graceful_registry
asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
@@ -55,9 +55,13 @@ 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]) :]
output_strs.append(recv_obj.decoded_texts[i] + new_text)
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 isinstance(recv_obj.finished_reason[i], FINISH_MATCHED_STR):
pos = output_strs[i].find(recv_obj.finished_reason[i].matched)
@@ -67,7 +71,8 @@ class DetokenizerManager:
self.send_to_tokenizer.send_pyobj(
BatchStrOut(
rids=recv_obj.rids,
output_str=output_strs,
output_strs=output_strs,
incomplete_strs=incomplete_strs,
meta_info=recv_obj.meta_info,
finished_reason=recv_obj.finished_reason,
)