"""LiteLLM Router as LangChain Model."""
from typing import (
Any,
AsyncIterator,
Iterator,
List,
Mapping,
Optional,
)
from langchain_core.callbacks.manager import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.language_models.chat_models import (
agenerate_from_stream,
generate_from_stream,
)
from langchain_core.messages import (
AIMessageChunk,
BaseMessage,
)
from langchain_core.outputs import (
ChatGeneration,
ChatGenerationChunk,
ChatResult,
)
from langchain_community.chat_models.litellm import (
ChatLiteLLM,
_convert_delta_to_message_chunk,
_convert_dict_to_message,
)
token_usage_key_name = "token_usage"
model_extra_key_name = "model_extra"
[docs]def get_llm_output(usage: Any, **params: Any) -> dict:
"""Get llm output from usage and params."""
llm_output = {token_usage_key_name: usage}
# copy over metadata (metadata came from router completion call)
metadata = params["metadata"]
for key in metadata:
if key not in llm_output:
# if token usage in metadata, prefer metadata's copy of it
llm_output[key] = metadata[key]
return llm_output
[docs]class ChatLiteLLMRouter(ChatLiteLLM):
"""LiteLLM Router as LangChain Model."""
router: Any
def __init__(self, *, router: Any, **kwargs: Any) -> None:
"""Construct Chat LiteLLM Router."""
super().__init__(**kwargs)
self.router = router
@property
def _llm_type(self) -> str:
return "LiteLLMRouter"
def _set_model_for_completion(self) -> None:
# use first model name (aka: model group),
# since we can only pass one to the router completion functions
self.model = self.router.model_list[0]["model_name"]
def _prepare_params_for_router(self, params: Any) -> None:
params["model"] = self.model
# allow the router to set api_base based on its model choice
api_base_key_name = "api_base"
if api_base_key_name in params and params[api_base_key_name] is None:
del params[api_base_key_name]
# add metadata so router can fill it below
params.setdefault("metadata", {})
def _generate(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
stream: Optional[bool] = None,
**kwargs: Any,
) -> ChatResult:
should_stream = stream if stream is not None else self.streaming
if should_stream:
stream_iter = self._stream(
messages, stop=stop, run_manager=run_manager, **kwargs
)
return generate_from_stream(stream_iter)
message_dicts, params = self._create_message_dicts(messages, stop)
params = {**params, **kwargs}
self._set_model_for_completion()
self._prepare_params_for_router(params)
response = self.router.completion(
messages=message_dicts,
**params,
)
return self._create_chat_result(response, **params)
def _stream(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Iterator[ChatGenerationChunk]:
default_chunk_class = AIMessageChunk
message_dicts, params = self._create_message_dicts(messages, stop)
params = {**params, **kwargs, "stream": True}
self._set_model_for_completion()
self._prepare_params_for_router(params)
for chunk in self.router.completion(messages=message_dicts, **params):
if len(chunk["choices"]) == 0:
continue
delta = chunk["choices"][0]["delta"]
chunk = _convert_delta_to_message_chunk(delta, default_chunk_class)
default_chunk_class = chunk.__class__
yield ChatGenerationChunk(message=chunk)
if run_manager:
run_manager.on_llm_new_token(chunk.content, **params)
async def _astream(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> AsyncIterator[ChatGenerationChunk]:
default_chunk_class = AIMessageChunk
message_dicts, params = self._create_message_dicts(messages, stop)
params = {**params, **kwargs, "stream": True}
self._set_model_for_completion()
self._prepare_params_for_router(params)
async for chunk in await self.router.acompletion(
messages=message_dicts, **params
):
if len(chunk["choices"]) == 0:
continue
delta = chunk["choices"][0]["delta"]
chunk = _convert_delta_to_message_chunk(delta, default_chunk_class)
default_chunk_class = chunk.__class__
yield ChatGenerationChunk(message=chunk)
if run_manager:
await run_manager.on_llm_new_token(chunk.content, **params)
async def _agenerate(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
stream: Optional[bool] = None,
**kwargs: Any,
) -> ChatResult:
should_stream = stream if stream is not None else self.streaming
if should_stream:
stream_iter = self._astream(
messages=messages, stop=stop, run_manager=run_manager, **kwargs
)
return await agenerate_from_stream(stream_iter)
message_dicts, params = self._create_message_dicts(messages, stop)
params = {**params, **kwargs}
self._set_model_for_completion()
self._prepare_params_for_router(params)
response = await self.router.acompletion(
messages=message_dicts,
**params,
)
return self._create_chat_result(response, **params)
# from
# https://github.com/langchain-ai/langchain/blob/master/libs/community/langchain_community/chat_models/openai.py
# but modified to handle LiteLLM Usage class
def _combine_llm_outputs(self, llm_outputs: List[Optional[dict]]) -> dict:
overall_token_usage: dict = {}
system_fingerprint = None
for output in llm_outputs:
if output is None:
# Happens in streaming
continue
token_usage = output["token_usage"]
if token_usage is not None:
# get dict from LiteLLM Usage class
for k, v in token_usage.dict().items():
if k in overall_token_usage:
overall_token_usage[k] += v
else:
overall_token_usage[k] = v
if system_fingerprint is None:
system_fingerprint = output.get("system_fingerprint")
combined = {"token_usage": overall_token_usage, "model_name": self.model_name}
if system_fingerprint:
combined["system_fingerprint"] = system_fingerprint
return combined
def _create_chat_result(
self, response: Mapping[str, Any], **params: Any
) -> ChatResult:
from litellm.utils import Usage
generations = []
for res in response["choices"]:
message = _convert_dict_to_message(res["message"])
gen = ChatGeneration(
message=message,
generation_info=dict(finish_reason=res.get("finish_reason")),
)
generations.append(gen)
token_usage = response.get("usage", Usage(prompt_tokens=0, total_tokens=0))
llm_output = get_llm_output(token_usage, **params)
return ChatResult(generations=generations, llm_output=llm_output)