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] Fixed when return_token_ids=False, the first event still contains prompt_token_ids. (#27561)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Oct 27, 2025 at 19:06 UTC
a4fc21895ed279930e5998eff4b9480da9d8442a
2 files changed
+13
-3
tests/entrypoints/openai/test_return_token_ids.py
+12
-2
@@ -27,8 +27,12 @@ def server():
27
28
29
@pytest.mark.asyncio
30
-async def test_basic_completion_with_emoji(server):
30
+@pytest.mark.parametrize("return_token_ids", [True, False, None])
31
+async def test_basic_completion_with_emoji(server, return_token_ids: bool | None):
32
"""Test basic completion with emoji to verify token_ids field."""
33
+ extra_body = None
34
+ if return_token_ids is not None:
35
+ extra_body = {"return_token_ids": return_token_ids}
36
async with server.get_async_client() as client:
37
# Test with return_token_ids enabled
38
completion = await client.completions.create(
@@ -37,7 +41,7 @@ async def test_basic_completion_with_emoji(server):
41
max_tokens=10,
42
temperature=0,
43
logprobs=1,
40
- extra_body={"return_token_ids": True},
44
+ extra_body=extra_body,
45
)
46
47
# Check the raw response to see the structure
@@ -45,6 +49,12 @@ async def test_basic_completion_with_emoji(server):
49
50
# Verify prompt_token_ids field is present in the completion response
51
assert "prompt_token_ids" in completion_dict["choices"][0]
52
+ if not return_token_ids:
53
+ # If return_token_ids is False, token_ids should not be present
54
+ assert completion_dict["choices"][0].get("token_ids") is None
55
+ assert completion_dict["choices"][0].get("prompt_token_ids") is None
56
+ # Skip further checks
57
+ return
58
assert isinstance(completion.choices[0].prompt_token_ids, list)
59
60
# Check against the expected prompt token IDs
vllm/entrypoints/openai/serving_completion.py
+1
-1
@@ -399,7 +399,7 @@ class OpenAIServingCompletion(OpenAIServing):
399
400
# has_echoed[i] is reused here to indicate whether
401
# we have already returned the prompt token IDs.
402
- if not has_echoed[i]:
402
+ if not has_echoed[i] and request.return_token_ids:
403
prompt_token_ids_to_return = prompt_token_ids
404
has_echoed[i] = True
405