What causes "不能正常加载MOSS的参数！" in binary-husky/gpt_academic?

Reproduce outside the thread: chdir to request_llms/moss and call the same moss_init logic to see the real traceback.

binary-husky/gpt_academic · critical · RuntimeError

不能正常加载MOSS的参数！

Error message

不能正常加载MOSS的参数！

What it means

Raised in the MOSS loader thread when self.moss_init() throws: the code first validate_path()s into request_llms/moss, then the bare except sends '[Local Message] Call MOSS fail' to the parent and raises RuntimeError('不能正常加载MOSS的参数！'). Root causes (missing weights, transformers version mismatch, CUDA OOM) are discarded.

Source

Thrown at request_llms/bridge_moss.py:119

        """
        self.prompt = self.meta_instruction
        self.local_history = []

    def run(self): # 子进程执行
        # 子进程执行
        # 第一次运行，加载参数
        def validate_path():
            import os, sys
            root_dir_assume = os.path.abspath(os.path.dirname(__file__) +  '/..')
            os.chdir(root_dir_assume + '/request_llms/moss')
            sys.path.append(root_dir_assume + '/request_llms/moss')
        validate_path() # validate path so you can run from base directory

        try:
            self.moss_init()
        except:
            self.child.send('[Local Message] Call MOSS fail 不能正常加载MOSS的参数。')
            raise RuntimeError("不能正常加载MOSS的参数！")

        # 进入任务等待状态
        # 这段代码来源 https://github.com/OpenLMLab/MOSS/blob/main/moss_cli_demo.py
        import torch
        while True:
            # 等待输入
            kwargs = self.child.recv()   # query = input("<|Human|>: ")
            try:
                query = kwargs['query']
                history = kwargs['history']
                sys_prompt = kwargs['sys_prompt']
                if len(self.local_history) > 0 and len(history)==0:
                    self.prompt = self.meta_instruction
                self.local_history.append(query)
                self.prompt += '<|Human|>: ' + query + '<eoh>'
                inputs = self.tokenizer(self.prompt, return_tensors="pt")
                with torch.no_grad():
                    outputs = self.model.generate(

View on GitHub (pinned to d6bde0fa54)

Solutions

Reproduce outside the thread: chdir to request_llms/moss and call the same moss_init logic to see the real traceback.
Fetch the MOSS weights/config exactly as request_llms/bridge_moss.py documents and pin the required transformers/torch versions.
Check GPU memory and LOCAL_MODEL_DEVICE; MOSS needs tens of GB VRAM (or quantization).
MOSS is unmaintained — prefer a maintained local backend if it cannot load.

Defensive patterns

Strategy: validation

Validate before calling

def moss_loadable() -> bool:
    import os
    return os.path.isdir(os.path.join(ROOT, 'request_llms', 'moss'))  # weights dir present
# gate the MOSS entry on moss_loadable() plus a dry import of moss_init deps

Try / catch

try:
    handle = GetGLMHandle()
except RuntimeError as e:
    if 'MOSS' in str(e):
        log.error('MOSS load failed: %s', e)

Prevention

Pre-download MOSS weights; pin transformers/torch versions
Verify VRAM fits MOSS before selecting it
Smoke-test moss_init standalone at startup

When it happens

Trigger: First use of the local MOSS model: moss_init fails to load the MOSS weights/config (missing moss-github checkout, failed download, transformers or torch version incompatible, insufficient GPU memory).

Common situations: Fresh clone without the MOSS model files, old transformers version incompatible with MOSS modeling code, LOCAL_MODEL_DEVICE=cuda on a CPU box, GPU with too little VRAM.

Related errors

AI-assisted analysis of binary-husky/gpt_academic@d6bde0fa54 (2026-08-14). Data as JSON: /api/errors/f83db06e4f596337. Report an issue: GitHub.