Very cool! In my company I had to serve BERT in rust using the tensorflow C API. If I knew of this framework I would have given it a shot, the examples are there and they look easy to understand.
I can’t believe I’ll get excited about phone specs again 🙌🏻✨. Do you think it could be possible to parallelize computation among various phones to run inference on transformer models? I assume is not worth it since you would need to transfer a ton of data among devices to run attention per layer, but the llama people have pulled so many tricks at this point…
Add comment