Some models bypass HF_ENDPOINT and download from huggingface.co
I set the environment variable HF_ENDPOINT="https://hf-mirror.com" to use a mirror for downloading models. While some models, like unsloth/Llama-3.2-3B-Instruct-bnb-4bit, correctly use the mirror, others, such as unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit, still attempt to download from huggingface.co.
Due to network restrictions, I cannot access huggingface.co. Even after manually downloading the model using huggingface-cli, the code still tries to connect to huggingface.co during model loading.
Is there a way to ensure all models respect the HF_ENDPOINT setting?
Interesting I'll take a look!
unsloth/Llama-3.2-3B-Instruct-bnb-4bit does not work as well
Hi, any updates on this issue? I’m facing the same problem.
Hi, any updates on this issue? I’m facing the same problem.
Hi, any updates on this issue? I’m facing the same problem.
+1