[Frontend] Expose logprob_token_ids on Python OpenAI endpoints (#43463)

Signed-off-by: Lang Zhao <lang.zhao@galileo.ai>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
langzhao-netizen
2026-07-13 14:40:21 -07:00
committed by GitHub
co-authored by Claude
parent 7738ef35b8
commit 8b8af2caf7
9 changed files with 401 additions and 7 deletions
@@ -568,13 +568,16 @@ class OpenAIServingChat(GenerateBaseServing):
if finish_reason_sent[i]:
continue
if request.logprobs and request.top_logprobs is not None:
if request.logprobs and (
request.top_logprobs is not None or request.logprob_token_ids
):
assert output.logprobs is not None, "Did not output logprobs"
logprobs = self._create_chat_logprobs(
token_ids=output.token_ids,
top_logprobs=output.logprobs,
tokenizer=tokenizer,
num_output_top_logprobs=request.top_logprobs,
logprob_token_ids=request.logprob_token_ids,
return_as_token_id=request.return_tokens_as_token_ids,
)
else:
@@ -862,12 +865,15 @@ class OpenAIServingChat(GenerateBaseServing):
token_ids = output.token_ids
out_logprobs = output.logprobs
if request.logprobs and request.top_logprobs is not None:
if request.logprobs and (
request.top_logprobs is not None or request.logprob_token_ids
):
assert out_logprobs is not None, "Did not output logprobs"
logprobs = self._create_chat_logprobs(
token_ids=token_ids,
top_logprobs=out_logprobs,
num_output_top_logprobs=request.top_logprobs,
logprob_token_ids=request.logprob_token_ids,
tokenizer=tokenizer,
return_as_token_id=request.return_tokens_as_token_ids,
)
@@ -1101,6 +1107,7 @@ class OpenAIServingChat(GenerateBaseServing):
top_logprobs: int | None,
tokenizer: TokenizerLike | None,
should_return_as_token_id: bool,
return_all: bool = False,
) -> list[ChatCompletionLogProb]:
return [
ChatCompletionLogProb(
@@ -1116,7 +1123,9 @@ class OpenAIServingChat(GenerateBaseServing):
bytes=list(token.encode("utf-8", errors="replace")),
)
for i, p in enumerate(logprobs.items())
if (top_logprobs and i < top_logprobs or top_logprobs == -1)
if return_all
or top_logprobs == -1
or (top_logprobs is not None and i < top_logprobs)
]
def _create_chat_logprobs(
@@ -1125,6 +1134,7 @@ class OpenAIServingChat(GenerateBaseServing):
top_logprobs: GenericSequence[dict[int, Logprob] | None],
tokenizer: TokenizerLike | None,
num_output_top_logprobs: int | None = None,
logprob_token_ids: list[int] | None = None,
return_as_token_id: bool | None = None,
) -> ChatCompletionLogProbs:
"""Create OpenAI-style logprobs."""
@@ -1177,6 +1187,7 @@ class OpenAIServingChat(GenerateBaseServing):
num_output_top_logprobs,
tokenizer,
should_return_as_token_id,
return_all=bool(logprob_token_ids),
),
)
)