[Bugfix] Map reasoning_effort to enable_thinking in chat template kwargs (#43401)

Signed-off-by: Ashwin Giridharan <girida@amazon.com>
Signed-off-by: Chauncey <chaunceyjiang@gmail.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
Ashwin Giridharan
2026-05-27 05:39:49 -07:00
committed by GitHub
co-authored by Chauncey
parent 2272062471
commit 52a31ccecc
4 changed files with 178 additions and 13 deletions
@@ -473,17 +473,27 @@ class ChatCompletionRequest(OpenAIBaseModel):
default_template: str | None,
default_template_content_format: ChatTemplateContentFormatOption,
) -> ChatParams:
extra_kwargs: dict[str, Any] = dict(
add_generation_prompt=self.add_generation_prompt,
continue_final_message=self.continue_final_message,
documents=self.documents,
reasoning_effort=self.reasoning_effort,
)
# When reasoning is requested, activate thinking for models whose
# chat templates require explicit opt-in (e.g., Gemma4 defaults
# enable_thinking to false). For templates that don't declare the
# variable, resolve_chat_template_kwargs filters it out harmlessly.
user_kwargs = self.chat_template_kwargs or {}
if self.reasoning_effort is not None and "enable_thinking" not in user_kwargs:
extra_kwargs["enable_thinking"] = self.reasoning_effort != "none"
return ChatParams(
chat_template=self.chat_template or default_template,
chat_template_content_format=default_template_content_format,
chat_template_kwargs=merge_kwargs(
self.chat_template_kwargs,
dict(
add_generation_prompt=self.add_generation_prompt,
continue_final_message=self.continue_final_message,
documents=self.documents,
reasoning_effort=self.reasoning_effort,
),
extra_kwargs,
),
media_io_kwargs=self.media_io_kwargs,
)
+17 -6
View File
@@ -298,17 +298,28 @@ class ResponsesRequest(OpenAIBaseModel):
continue_final = should_continue_final_message(self.input)
reasoning = self.reasoning
reasoning_effort = None if reasoning is None else reasoning.effort
extra_kwargs: dict[str, Any] = dict(
add_generation_prompt=not continue_final,
continue_final_message=continue_final,
reasoning_effort=reasoning_effort,
)
# When reasoning is requested, activate thinking for models whose
# chat templates require explicit opt-in (e.g., Gemma4 defaults
# enable_thinking to false). For templates that don't declare the
# variable, resolve_chat_template_kwargs filters it out harmlessly.
user_kwargs = self.chat_template_kwargs or {}
if reasoning_effort is not None and "enable_thinking" not in user_kwargs:
extra_kwargs["enable_thinking"] = reasoning_effort != "none"
return ChatParams(
chat_template=default_template,
chat_template_content_format=default_template_content_format,
chat_template_kwargs=merge_kwargs( # To remove unset values
chat_template_kwargs=merge_kwargs(
self.chat_template_kwargs,
dict(
add_generation_prompt=not continue_final,
continue_final_message=continue_final,
reasoning_effort=None if reasoning is None else reasoning.effort,
),
extra_kwargs,
),
media_io_kwargs=self.media_io_kwargs,
)