phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-10-12 13:57:57 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

featu: support passing "extra body" throught to providers

Browse source 
# What does this PR do?
Allows passing through extra_body parameters to inference providers.


closes #2720

## Test Plan
CI and added new test
This commit is contained in:
Eric Huang 2025-10-10 14:05:16 -07:00
parent cb7fb0705b
commit d7b57a8dd2
4 changed files with 107 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

10
llama_stack/providers/utils/inference/litellm_openai_mixin.py
View file

@ -230,6 +230,9 @@ class LiteLLMOpenAIMixin(
) -> OpenAICompletion:
model_obj = await self.model_store.get_model(params.model)
# Extract extra fields
extra_body = dict(params.__pydantic_extra__ or {})
request_params = await prepare_openai_completion_params(
model=self.get_litellm_model_name(model_obj.provider_resource_id),
prompt=params.prompt,
@ -248,11 +251,10 @@ class LiteLLMOpenAIMixin(
temperature=params.temperature,
top_p=params.top_p,
user=params.user,
guided_choice=params.guided_choice,
prompt_logprobs=params.prompt_logprobs,
suffix=params.suffix,
api_key=self.get_api_key(),
api_base=self.api_base,
**extra_body,
)
return await litellm.atext_completion(**request_params)
@ -272,6 +274,9 @@ class LiteLLMOpenAIMixin(
model_obj = await self.model_store.get_model(params.model)
# Extract extra fields
extra_body = dict(params.__pydantic_extra__ or {})
request_params = await prepare_openai_completion_params(
model=self.get_litellm_model_name(model_obj.provider_resource_id),
messages=params.messages,
@ -298,6 +303,7 @@ class LiteLLMOpenAIMixin(
user=params.user,
api_key=self.get_api_key(),
api_base=self.api_base,
**extra_body,
)
return await litellm.acompletion(**request_params)

15
llama_stack/providers/utils/inference/openai_mixin.py
View file

@ -228,15 +228,6 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
"""
Direct OpenAI completion API call.
"""
# Handle parameters that are not supported by OpenAI API, but may be by the provider
# prompt_logprobs is supported by vLLM
# guided_choice is supported by vLLM
# TODO: test coverage
extra_body: dict[str, Any] = {}
if params.prompt_logprobs is not None and params.prompt_logprobs >= 0:
extra_body["prompt_logprobs"] = params.prompt_logprobs
if params.guided_choice:
extra_body["guided_choice"] = params.guided_choice
# TODO: fix openai_completion to return type compatible with OpenAI's API response
completion_kwargs = await prepare_openai_completion_params(
@ -259,7 +250,9 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
user=params.user,
suffix=params.suffix,
)
resp = await self.client.completions.create(**completion_kwargs, extra_body=extra_body)
if extra_body := dict(params.__pydantic_extra__ or {}):
completion_kwargs["extra_body"] = extra_body
resp = await self.client.completions.create(**completion_kwargs)
return await self._maybe_overwrite_id(resp, params.stream) # type: ignore[no-any-return]
@ -316,6 +309,8 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
user=params.user,
)
if extra_body := dict(params.__pydantic_extra__ or {}):
request_params["extra_body"] = extra_body
resp = await self.client.chat.completions.create(**request_params)
return await self._maybe_overwrite_id(resp, params.stream) # type: ignore[no-any-return]