Fosowl/agenticSeek · critical · Exception

Failed to load the routing model. Please run the dl_safetens

Error message

Failed to load the routing model. Please run the dl_safetensors.sh script inside llm_router/ directory to download the model.

What it means

load_llm_router loads the AdaptiveClassifier safetensors model from ./llm_router (relative to CWD). If loading fails for any reason it raises a single Exception instructing the developer to run dl_safetensors.sh inside llm_router/ to download the model weights. The original exception is swallowed, so the message always points at the most common cause: missing weights.

Source

Thrown at sources/router.py:58

        animate_thinking("Loading zero-shot pipeline...", color="status")
        return {
            "bart": pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
        }

    def load_llm_router(self) -> AdaptiveClassifier:
        """
        Load the LLM router model.
        returns:
            AdaptiveClassifier: The loaded model
        exceptions:
            Exception: If the safetensors fails to load
        """
        path = "../llm_router" if __name__ == "__main__" else "./llm_router"
        try:
            animate_thinking("Loading LLM router model...", color="status")
            talk_classifier = AdaptiveClassifier.from_pretrained(path)
        except Exception as e:
            raise Exception("Failed to load the routing model. Please run the dl_safetensors.sh script inside llm_router/ directory to download the model.")
        return talk_classifier

    def get_device(self) -> str:
        if torch.backends.mps.is_available():
            return "mps"
        elif torch.cuda.is_available():
            return "cuda:0"
        else:
            return "cpu"
    
    def learn_few_shots_complexity(self) -> None:
        """
        Few shot learning for complexity estimation.
        Use the build in add_examples method of the Adaptive_classifier.
        """
        few_shots = [
            ("hi", "LOW"),
            ("How it's going ?", "LOW"),

View on GitHub (pinned to ae57a23577)

Solutions

1. Run `bash dl_safetensors.sh` inside the llm_router/ directory to download the model
2. Launch the app from the repository root so ./llm_router resolves correctly (or fix the relative path in router.py:53 to be __file__-based)
3. Verify llm_router/ contains config.json and the .safetensors files; re-download if corrupt
4. Check disk space and HuggingFace connectivity if the download script itself fails

Example fix

// before
path = "../llm_router" if __name__ == "__main__" else "./llm_router"
// after
import os
base = os.path.dirname(os.path.abspath(__file__))
path = os.path.join(os.path.dirname(base), "llm_router") if __name__ == "__main__" else os.path.join(base, "llm_router")

import os
path = './llm_router'
weights = [f for f in (os.listdir(path) if os.path.isdir(path) else []) if f.endswith('.safetensors')]
if not weights:
    raise SystemExit('Model weights missing. Run: bash llm_router/dl_safetensors.sh')

try:
    router = Router()
except Exception as e:
    if 'dl_safetensors.sh' in str(e):
        import subprocess
        subprocess.run(['bash', 'llm_router/dl_safetensors.sh'], check=True)
        router = Router()
    else:
        raise

When it happens

Trigger: Instantiating the router (Router.__init__ -> load_llm_router) when AdaptiveClassifier.from_pretrained('./llm_router') fails — directory absent, weights never downloaded, partial/corrupt download, or CWD is not the repo root so the relative path resolves wrong.

Common situations: Fresh clone without running dl_safetensors.sh; launching the app from a different working directory (the path is CWD-relative, not file-relative); interrupted download leaving corrupt safetensors; missing config.json alongside weights.

Related errors

• Prompt file not found at path: {file_path}
• Could not find: {path}

AI-assisted analysis of Fosowl/agenticSeek@ae57a23577 (2026-08-30). Data as JSON: /api/errors/ad13c7ebccd375ca. Report an issue: GitHub.