text-generation-inference
text-generation-inference copied to clipboard
Changing convert logic.
Should be more robust to shared tensors (ok when using
`from_pretrained). But forcing us to add new checks in our loading
code (since the chosen key to keep might be different from
`transformers`).