68 lines
2.8 KiB
Python
68 lines
2.8 KiB
Python
# coding=utf-8
|
||
"""
|
||
@project: maxkb
|
||
@Author:虎
|
||
@file: llm.py
|
||
@date:2024/3/6 11:48
|
||
@desc:
|
||
"""
|
||
from typing import List, Dict
|
||
from urllib.parse import urlparse, ParseResult
|
||
|
||
from langchain_core.messages import BaseMessage, get_buffer_string
|
||
from langchain_ollama.chat_models import ChatOllama
|
||
|
||
|
||
from common.config.tokenizer_manage_config import TokenizerManage
|
||
from models_provider.base_model_provider import MaxKBBaseModel
|
||
|
||
|
||
def get_base_url(url: str):
|
||
parse = urlparse(url)
|
||
result_url = ParseResult(scheme=parse.scheme, netloc=parse.netloc, path=parse.path, params='',
|
||
query='',
|
||
fragment='').geturl()
|
||
return result_url[:-1] if result_url.endswith("/") else result_url
|
||
|
||
|
||
class OllamaChatModel(MaxKBBaseModel, ChatOllama):
|
||
model_resource_id: str | None = None
|
||
@staticmethod
|
||
def is_cache_model():
|
||
return False
|
||
|
||
@staticmethod
|
||
def new_instance(model_type, model_name, model_credential: Dict[str, object], **model_kwargs):
|
||
api_base = model_credential.get('api_base', '')
|
||
base_url = get_base_url(api_base)
|
||
optional_params = MaxKBBaseModel.filter_optional_params(model_kwargs)
|
||
from models_provider.context_policy import model_context
|
||
optional_params['num_ctx'] = model_context(model_kwargs)
|
||
|
||
return OllamaChatModel(model=model_name, base_url=base_url,
|
||
model_resource_id=str(model_kwargs['model_id']) if model_kwargs.get('model_id') else None,
|
||
stream=True, **optional_params)
|
||
|
||
def _create_chat_stream(self, messages, stop=None, **kwargs):
|
||
from common.model_runtime_client import inference_lease
|
||
with inference_lease(self.model_resource_id) as state:
|
||
if state:
|
||
kwargs['keep_alive'] = -1 if state['strategy'] == 'PRELOAD' else state['idle_seconds']
|
||
yield from super()._create_chat_stream(messages, stop, **kwargs)
|
||
|
||
async def _acreate_chat_stream(self, messages, stop=None, **kwargs):
|
||
from common.model_runtime_client import async_inference_lease
|
||
async with async_inference_lease(self.model_resource_id) as state:
|
||
if state:
|
||
kwargs['keep_alive'] = -1 if state['strategy'] == 'PRELOAD' else state['idle_seconds']
|
||
async for part in super()._acreate_chat_stream(messages, stop, **kwargs):
|
||
yield part
|
||
|
||
def get_num_tokens_from_messages(self, messages: List[BaseMessage]) -> int:
|
||
tokenizer = TokenizerManage.get_tokenizer()
|
||
return sum([len(tokenizer.encode(get_buffer_string([m]))) for m in messages])
|
||
|
||
def get_num_tokens(self, text: str) -> int:
|
||
tokenizer = TokenizerManage.get_tokenizer()
|
||
return len(tokenizer.encode(text))
|