support returning tokenids in responses api (#33212)

Signed-off-by: Christian Munley <cmunley@nvidia.com>
This commit is contained in:
cmunley1
2026-01-29 16:52:39 +08:00
committed by GitHub
parent 53fc166402
commit 3bba2edb0f
3 changed files with 50 additions and 9 deletions
+1 -1
View File
@@ -1528,7 +1528,7 @@ class OpenAIServing:
"Unable to get tokenizer because `skip_tokenizer_init=True`"
)
return tokenizer.decode(token_id)
return tokenizer.decode([token_id])
def _is_model_supported(self, model_name: str | None) -> bool:
if not model_name:
+10 -8
View File
@@ -836,10 +836,11 @@ class OpenAIServingResponses(OpenAIServing):
for i, (token_id, _logprob) in enumerate(logprobs.items()):
if i >= top_logprobs:
break
text = (
_logprob.decoded_token
if _logprob.decoded_token is not None
else tokenizer.decode([token_id])
text = self._get_decoded_token(
logprob=_logprob,
token_id=token_id,
tokenizer=tokenizer,
return_as_token_id=self.return_tokens_as_token_ids,
)
out.append(
LogprobTopLogprob(
@@ -865,10 +866,11 @@ class OpenAIServingResponses(OpenAIServing):
for i, token_id in enumerate(token_ids):
logprob = logprobs[i]
token_logprob = logprob[token_id]
text = (
token_logprob.decoded_token
if token_logprob.decoded_token is not None
else tokenizer.decode([token_id])
text = self._get_decoded_token(
logprob=token_logprob,
token_id=token_id,
tokenizer=tokenizer,
return_as_token_id=self.return_tokens_as_token_ids,
)
out.append(
Logprob(