my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

#1	import json
#2	import os
#3	from typing import Dict, List, Optional, Union
#4
#5	from openai import OpenAI
#6
#7	from mem0.configs.llms.base import BaseLlmConfig
#8	from mem0.configs.llms.vllm import VllmConfig
#9	from mem0.llms.base import LLMBase
#10	from mem0.memory.utils import extract_json
#11
#12
#13	class VllmLLM(LLMBase):
#14	def __init__(self, config: Optional[Union[BaseLlmConfig, VllmConfig, Dict]] = None):
#15	# Convert to VllmConfig if needed
#16	if config is None:
#17	config = VllmConfig()
#18	elif isinstance(config, dict):
#19	config = VllmConfig(**config)
#20	elif isinstance(config, BaseLlmConfig) and not isinstance(config, VllmConfig):
#21	# Convert BaseLlmConfig to VllmConfig
#22	config = VllmConfig(
#23	model=config.model,
#24	temperature=config.temperature,
#25	api_key=config.api_key,
#26	max_tokens=config.max_tokens,
#27	top_p=config.top_p,
#28	top_k=config.top_k,
#29	enable_vision=config.enable_vision,
#30	vision_details=config.vision_details,
#31	http_client_proxies=config.http_client,
#32	)
#33
#34	super().__init__(config)
#35
#36	if not self.config.model:
#37	self.config.model = "Qwen/Qwen2.5-32B-Instruct"
#38
#39	self.config.api_key = self.config.api_key or os.getenv("VLLM_API_KEY") or "vllm-api-key"
#40	base_url = self.config.vllm_base_url or os.getenv("VLLM_BASE_URL")
#41	self.client = OpenAI(api_key=self.config.api_key, base_url=base_url)
#42
#43	def _parse_response(self, response, tools):
#44	"""
#45	Process the response based on whether tools are used or not.
#46
#47	Args:
#48	response: The raw response from API.
#49	tools: The list of tools provided in the request.
#50
#51	Returns:
#52	str or dict: The processed response.
#53	"""
#54	if tools:
#55	processed_response = {
#56	"content": response.choices[0].message.content,
#57	"tool_calls": [],
#58	}
#59
#60	if response.choices[0].message.tool_calls:
#61	for tool_call in response.choices[0].message.tool_calls:
#62	processed_response["tool_calls"].append(
#63	{
#64	"name": tool_call.function.name,
#65	"arguments": json.loads(extract_json(tool_call.function.arguments)),
#66	}
#67	)
#68
#69	return processed_response
#70	else:
#71	return response.choices[0].message.content
#72
#73	def generate_response(
#74	self,
#75	messages: List[Dict[str, str]],
#76	response_format=None,
#77	tools: Optional[List[Dict]] = None,
#78	tool_choice: str = "auto",
#79	**kwargs,
#80	):
#81	"""
#82	Generate a response based on the given messages using vLLM.
#83
#84	Args:
#85	messages (list): List of message dicts containing 'role' and 'content'.
#86	response_format (str or object, optional): Format of the response. Defaults to "text".
#87	tools (list, optional): List of tools that the model can call. Defaults to None.
#88	tool_choice (str, optional): Tool choice method. Defaults to "auto".
#89	**kwargs: Additional vLLM-specific parameters.
#90
#91	Returns:
#92	str: The generated response.
#93	"""
#94	params = self._get_supported_params(messages=messages, **kwargs)
#95	params.update(
#96	{
#97	"model": self.config.model,
#98	"messages": messages,
#99	}
#100	)
#101
#102	if tools:
#103	params["tools"] = tools
#104	params["tool_choice"] = tool_choice
#105
#106	response = self.client.chat.completions.create(**params)
#107	return self._parse_response(response, tools)
#108

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public