[Frontend] Responses API messages out, just harmony for now (#24985)

Signed-off-by: Alec Solder <alecs@fb.com>
Co-authored-by: Alec Solder <alecs@fb.com>
Co-authored-by: Ye (Charlotte) Qi <yeq@meta.com>
This commit is contained in:
Alec S
2025-09-19 21:40:16 +00:00
committed by GitHub
co-authored by Alec Solder Ye Qi
parent ddc9048394
commit e69e0b8b5f
3 changed files with 44 additions and 1 deletions
@@ -475,9 +475,14 @@ class OpenAIServingResponses(OpenAIServing):
# "completed" is implemented as the "catch-all" for now.
status: ResponseStatus = "completed"
input_messages = None
output_messages = None
if self.use_harmony:
assert isinstance(context, HarmonyContext)
output = self._make_response_output_items_with_harmony(context)
if request.enable_response_messages:
input_messages = context.messages[:context.num_init_messages]
output_messages = context.messages[context.num_init_messages:]
num_tool_output_tokens = context.num_tool_output_tokens
if len(output) > 0:
if context.finish_reason == "length":
@@ -496,6 +501,12 @@ class OpenAIServingResponses(OpenAIServing):
output = self._make_response_output_items(request, final_output,
tokenizer)
# TODO: context for non-gptoss models doesn't use messages
# so we can't get them out yet
if request.enable_response_messages:
raise NotImplementedError(
"enable_response_messages is currently"
" only supported for gpt-oss")
# Calculate usage.
assert final_res.prompt_token_ids is not None
num_tool_output_tokens = 0
@@ -519,6 +530,8 @@ class OpenAIServingResponses(OpenAIServing):
response = ResponsesResponse.from_request(
request,
sampling_params,
input_messages=input_messages,
output_messages=output_messages,
model_name=model_name,
created_time=created_time,
output=output,