A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 01:15 UTC
[gpt-oss] Fix harmony parser in streaming responses (#30205)
Signed-off-by: AlonKejzman <alonkeizman@gmail.com>
AlonKejzman committed
Dec 22, 2025 at 14:56 UTC
bd6d5a7475aa26a85de5d63cfcae9a1151d7652f
1 file changed
+5
vllm/entrypoints/openai/serving_chat.py
+5
@@ -811,6 +811,11 @@ class OpenAIServingChat(OpenAIServing):
811
delta_text += harmony_parser.last_content_delta or ""
812
cur_channel = harmony_parser.current_channel
813
cur_recipient = harmony_parser.current_recipient
814
+ # handle the case where several tokens where generated at once
815
+ # including the final token, leading to a delta in the text
816
+ # but the current channel to be empty (start state)
817
+ if not cur_channel and delta_text:
818
+ cur_channel = "final"
819
else:
820
delta_text = output.text
821