​
SGLangModel

class SGLangModel(BaseModelBackend):
SGLang service interface. Parameters:
  • model_type (Union[ModelType, str]): Model for which a backend is created.
  • model_config_dict (Optional[Dict[str, Any]], optional): A dictionary that will be fed into:obj:openai.ChatCompletion.create(). If :obj:None, :obj:SGLangConfig().as_dict() will be used. (default: :obj:None)
  • api_key (Optional[str], optional): The API key for authenticating with the model service. SGLang doesn’t need API key, it would be ignored if set. (default: :obj:None)
  • url (Optional[str], optional): The url to the model service. If not provided, :obj:"http://127.0.0.1:30000/v1" will be used. (default: :obj:None)
  • token_counter (Optional[BaseTokenCounter], optional): Token counter to use for the model. If not provided, :obj:OpenAITokenCounter( ModelType.GPT_4O_MINI) will be used. (default: :obj:None)
  • timeout (Optional[float], optional): The timeout value in seconds for API calls. If not provided, will fall back to the MODEL_TIMEOUT environment variable or default to 180 seconds. (default: :obj:None)
  • max_retries (int, optional): Maximum number of retries for API calls. (default: :obj:3) **kwargs (Any): Additional arguments to pass to the client initialization.
  • Reference: https://sgl-project.github.io/backend/openai_api_completions. html

​
init

def __init__(
    self,
    model_type: Union[ModelType, str],
    model_config_dict: Optional[Dict[str, Any]] = None,
    api_key: Optional[str] = None,
    url: Optional[str] = None,
    token_counter: Optional[BaseTokenCounter] = None,
    timeout: Optional[float] = None,
    max_retries: int = 3,
    **kwargs: Any
):

​
_start_server

def _start_server(self):

​
_ensure_server_running

def _ensure_server_running(self):
Ensures that the server is running. If not, starts the server.

​
_monitor_inactivity

def _monitor_inactivity(self):
Monitor whether the server process has been inactive for over 10 minutes.

​
token_counter

def token_counter(self):
Returns: BaseTokenCounter: The token counter following the model’s tokenization style.

​
_run

def _run(
    self,
    messages: List[OpenAIMessage],
    response_format: Optional[Type[BaseModel]] = None,
    tools: Optional[List[Dict[str, Any]]] = None
):
Runs inference of OpenAI chat completion. Parameters:
  • messages (List[OpenAIMessage]): Message list with the chat history in OpenAI API format.
Returns: Union[ChatCompletion, Stream[ChatCompletionChunk]]: ChatCompletion in the non-stream mode, or Stream[ChatCompletionChunk] in the stream mode.

​
stream

def stream(self):
Returns: bool: Whether the model is in stream mode.

​
del

def __del__(self):
Properly clean up resources when the model is destroyed.

​
cleanup

def cleanup(self):
Terminate the server process and clean up resources.

​
_terminate_process

def _terminate_process(process):

​
_kill_process_tree

def _kill_process_tree(
    parent_pid,
    include_parent: bool = True,
    skip_pid: Optional[int] = None
):
Kill the process and all its child processes.

​
_execute_shell_command

def _execute_shell_command(command: str):
Execute a shell command and return the process handle Parameters:
  • command: Shell command as a string (can include \ line continuations)
Returns: subprocess.Popen: Process handle

​
_wait_for_server

def _wait_for_server(base_url: str, timeout: Optional[float] = 30):
Wait for the server to be ready by polling the /v1/models endpoint. Parameters:
  • base_url (str): The base URL of the server
  • timeout (Optional[float]): Maximum time to wait in seconds. (default: :obj:30)