binary-husky/gpt_academic · error · RuntimeError

_llm_handle.get_state()

Error message

_llm_handle.get_state()

What it means

In predict_no_ui_long_connection (the non-UI plugin pathway), after obtaining the singleton local-LLM handle it checks _llm_handle.running; if the loader subprocess died (e.g. it raised error 182's RuntimeError), running is False and this raises RuntimeError with the handle's last state string (e.g. '`加载模型失败`'). The message is the state label, not a stack trace.

Source

Thrown at request_llms/local_llm_class.py:227

                    self.running = False
                    self.corrupted = True
                    break
                else:
                    std_out = ""
                    yield res

def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='classic'):
    load_message = f"{model_name}尚未加载，加载需要一段时间。注意，取决于`config.py`的配置，{model_name}消耗大量的内存（CPU）或显存（GPU），也许会导致低配计算机卡死 ……"

    def predict_no_ui_long_connection(inputs:str, llm_kwargs:dict, history:list=[], sys_prompt:str="", observe_window:list=[], console_silence:bool=False):
        """
            refer to request_llms/bridge_all.py
        """
        _llm_handle = GetSingletonHandle().get_llm_model_instance(LLMSingletonClass)
        if len(observe_window) >= 1:
            observe_window[0] = load_message + "\n\n" + _llm_handle.get_state()
        if not _llm_handle.running:
            raise RuntimeError(_llm_handle.get_state())

        if history_format == 'classic':
            # 没有 sys_prompt 接口，因此把prompt加入 history
            history_feedin = []
            history_feedin.append([sys_prompt, "Certainly!"])
            for i in range(len(history)//2):
                history_feedin.append([history[2*i], history[2*i+1]])
        elif history_format == 'chatglm3':
            # 有 sys_prompt 接口
            conversation_cnt = len(history) // 2
            history_feedin = [{"role": "system", "content": sys_prompt}]
            if conversation_cnt:
                for index in range(0, 2*conversation_cnt, 2):
                    what_i_have_asked = {}
                    what_i_have_asked["role"] = "user"
                    what_i_have_asked["content"] = history[index]
                    what_gpt_answer = {}
                    what_gpt_answer["role"] = "assistant"

View on GitHub (pinned to d6bde0fa54)

Solutions

  1. Reproduce with the interactive predict() path first: its chatbot output shows the full '[Local Message]' traceback from the loader.
  2. Fix the underlying load failure (model path, memory, dependencies) — see error 182.
  3. Restart the app to rebuild the singleton handle after fixing config; a dead handle never recovers in-process.
  4. If memory-bound, choose a smaller model or add swap/quantization.
Defensive patterns

Strategy: validation

Validate before calling

_llm_handle = GetSingletonHandle().get_llm_model_instance(LLMSingletonClass)
if not getattr(_llm_handle, 'running', False):
    # fail fast with the loader's last state instead of proceeding
    raise ModelNotReady(_llm_handle.get_state())

Try / catch

try:
    result = predict_no_ui_long_connection(...)
except RuntimeError as e:
    if '加载模型失败' in str(e) or '失败' in str(e):
        restart_model_subprocess_with_diagnostics()

Prevention

When it happens

Trigger: Calling predict_no_ui_long_connection for a local model whose background load failed or crashed: OOM kill of the loader process, exception during load_model_and_tokenizer, or the singleton was constructed moments ago and died before this call.

Common situations: Plugins/crazy_functions that use the no-ui pipeline against a local model on a machine with insufficient memory; first request after configuring a local model that fails to download; handle poisoned by a previous failed run.

Related errors

AI-assisted analysis of binary-husky/gpt_academic@d6bde0fa54 (2026-08-14). Data as JSON: /api/errors/726b05af2f580690. Report an issue: GitHub.