reduce Llama 3 70B from 4 to 2 H100s

[philipkiely-baseten]

Llama 3 70B only needs 2, not 4, 80 GB GPUs to run inference. I've changed the config and tested the updated config in production.