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 14, 05:30 UTC
[Bugfix] Allow `--skip-tokenizer-init` with `echo and return_token_ids` (#26238)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Oct 5, 2025 at 13:38 UTC
a964e5e6c35e8f22bd7663dcf93d1c801421a029
2 files changed
+3
-2
tests/entrypoints/openai/test_token_in_token_out.py
+1
-1
@@ -54,7 +54,7 @@ async def test_token_in_token_out_and_logprobs(server):
54
prompt=token_ids,
55
max_tokens=20,
56
temperature=0,
57
- echo=False,
57
+ echo=True,
58
extra_body={
59
"return_token_ids": True,
60
},
vllm/entrypoints/openai/serving_completion.py
+2
-1
@@ -691,5 +691,6 @@ class OpenAIServingCompletion(OpenAIServing):
691
truncate_prompt_tokens=request.truncate_prompt_tokens,
692
add_special_tokens=request.add_special_tokens,
693
cache_salt=request.cache_salt,
694
- needs_detokenization=bool(request.echo),
694
+ needs_detokenization=bool(request.echo
695
+ and not request.return_token_ids),
696
)