108 lines
3.7 KiB
Python
108 lines
3.7 KiB
Python
import json
|
|
import os
|
|
from typing import Dict, List, Optional, Union
|
|
|
|
from openai import OpenAI
|
|
|
|
from neomem.configs.llms.base import BaseLlmConfig
|
|
from neomem.configs.llms.vllm import VllmConfig
|
|
from neomem.llms.base import LLMBase
|
|
from neomem.memory.utils import extract_json
|
|
|
|
|
|
class VllmLLM(LLMBase):
|
|
def __init__(self, config: Optional[Union[BaseLlmConfig, VllmConfig, Dict]] = None):
|
|
# Convert to VllmConfig if needed
|
|
if config is None:
|
|
config = VllmConfig()
|
|
elif isinstance(config, dict):
|
|
config = VllmConfig(**config)
|
|
elif isinstance(config, BaseLlmConfig) and not isinstance(config, VllmConfig):
|
|
# Convert BaseLlmConfig to VllmConfig
|
|
config = VllmConfig(
|
|
model=config.model,
|
|
temperature=config.temperature,
|
|
api_key=config.api_key,
|
|
max_tokens=config.max_tokens,
|
|
top_p=config.top_p,
|
|
top_k=config.top_k,
|
|
enable_vision=config.enable_vision,
|
|
vision_details=config.vision_details,
|
|
http_client_proxies=config.http_client,
|
|
)
|
|
|
|
super().__init__(config)
|
|
|
|
if not self.config.model:
|
|
self.config.model = "Qwen/Qwen2.5-32B-Instruct"
|
|
|
|
self.config.api_key = self.config.api_key or os.getenv("VLLM_API_KEY") or "vllm-api-key"
|
|
base_url = self.config.vllm_base_url or os.getenv("VLLM_BASE_URL")
|
|
self.client = OpenAI(api_key=self.config.api_key, base_url=base_url)
|
|
|
|
def _parse_response(self, response, tools):
|
|
"""
|
|
Process the response based on whether tools are used or not.
|
|
|
|
Args:
|
|
response: The raw response from API.
|
|
tools: The list of tools provided in the request.
|
|
|
|
Returns:
|
|
str or dict: The processed response.
|
|
"""
|
|
if tools:
|
|
processed_response = {
|
|
"content": response.choices[0].message.content,
|
|
"tool_calls": [],
|
|
}
|
|
|
|
if response.choices[0].message.tool_calls:
|
|
for tool_call in response.choices[0].message.tool_calls:
|
|
processed_response["tool_calls"].append(
|
|
{
|
|
"name": tool_call.function.name,
|
|
"arguments": json.loads(extract_json(tool_call.function.arguments)),
|
|
}
|
|
)
|
|
|
|
return processed_response
|
|
else:
|
|
return response.choices[0].message.content
|
|
|
|
def generate_response(
|
|
self,
|
|
messages: List[Dict[str, str]],
|
|
response_format=None,
|
|
tools: Optional[List[Dict]] = None,
|
|
tool_choice: str = "auto",
|
|
**kwargs,
|
|
):
|
|
"""
|
|
Generate a response based on the given messages using vLLM.
|
|
|
|
Args:
|
|
messages (list): List of message dicts containing 'role' and 'content'.
|
|
response_format (str or object, optional): Format of the response. Defaults to "text".
|
|
tools (list, optional): List of tools that the model can call. Defaults to None.
|
|
tool_choice (str, optional): Tool choice method. Defaults to "auto".
|
|
**kwargs: Additional vLLM-specific parameters.
|
|
|
|
Returns:
|
|
str: The generated response.
|
|
"""
|
|
params = self._get_supported_params(messages=messages, **kwargs)
|
|
params.update(
|
|
{
|
|
"model": self.config.model,
|
|
"messages": messages,
|
|
}
|
|
)
|
|
|
|
if tools:
|
|
params["tools"] = tools
|
|
params["tool_choice"] = tool_choice
|
|
|
|
response = self.client.chat.completions.create(**params)
|
|
return self._parse_response(response, tools)
|