Is the Grok-2 model supported?

[edyapd]

If Grok-2 is supported, how can I run it?
Currently, I'm getting the following error on load:
llama_model_load: error loading model: error loading model vocabulary: unknown pre-tokenizer type: 'grok-2'

[firecoperana]

#782

It's supported now. You can build from this PR or wait after it's merged to main.

[edyapd]

Thank you so much.
Yes, Grok-2 is working now. I tried it with the grok-2-UD-Q2_K_XL model.
Unfortunately, I'm only able to load 45/65 layers onto my Tesla P40. As a result, the inference is very slow, around 0.9 tokens/second.

[firecoperana]

Yes, it's a very large model since it has at least 100b active parameters.