-
Notifications
You must be signed in to change notification settings - Fork 17
Description
I am attempting to run Nemotron4-340b testing on a single GB300 rack configuration (64 GPU), bf16. I am running into some issues in regards to the latest build (25.10). This build added HuggingFace authentication as a requirement for Nemotron4-340b, but I am unable to get my authentication to work.
ValueError: Unable to instantiate HuggingFace AUTOTOKENIZER for nvidia/Nemotron-4-340B-Base. Exception: We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files.
Check your internet connection or see how to run the library in offline mode at 'https://huggingface.co/docs/transformers/installation#offline-mode'. [huggingface.co]
Fiddle context: failed to construct or call AutoTokenizer at .tokenizer with positional arguments: (), keyword arguments: (pretrained_model_name='nvidia/Nemotron-4-340B-Base', use_fast=True).
I have created a new account and created a token with write permissions, and provided it to the install. I then proceeded to add the HF_TOKEN parameter in manually through launch.sh, and also logged in through each virtual environment. However, I am still running into this error.
I have been able to ping the huggingface IP address (3.170.185.35), as well as been able to run a curl command using the token. However, the script fails to connect.
I have reverted to the 25.08 build, which works fine for me. Do you know what could be causing this authentication error?