Compatibility with `hf download`

[Adel-Moumen]

Hi guys,

Awesome library, really underrated! Could I ask whether there’s a way to download kernels via hf download? So far, I’ve been relying on the kernels toolkit by running some Python code, but I think this library would integrate nicely with hf download.

Also, besides attn_implementation in transformers, are you aware of other ways to monkey-patch pre-trained Transformers (e.g., Llama) to use fast CUDA kernels (e.g., activations etc) from kernels ?

Thanks!

[sayakpaul]

Could I ask whether there’s a way to download kernels via hf download?

What happens when you do hf download kernels-community/vllm-flash-attn3?

are you aware of other ways to monkey-patch pre-trained Transformers (e.g., Llama) to use fast CUDA kernels (e.g., activations etc) from kernels ?

https://www.youtube.com/watch?v=q0GfzJmuaUM might be helpful in this regard.