[don't merge] Branch for AMD GPUs (with older bitsandbytes)

[borzunov]

See the "Running on AMD GPUs" tutorial to learn how to use this branch.

This branch supports bitsandbytes < 0.40.0, so that AMD GPUs owners can use bitandbytes-rocm ports from bitsandbytes-foundation/bitsandbytes#47

To make it possible, we remove features depending on bitsandbytes >= 0.40.0 (NF4 and LoRA adapters). In case of the LoRA adapters, they depend on the peft library that requires bitsandbytes >= 0.40.0.

[PaulConyngham]

would be awesome to see this integrated

[borzunov]

Hi @PaulConyngham,

You already can run a server on AMD GPUs by following this tutorial: https://github.com/bigscience-workshop/petals/wiki/Running-on-AMD-GPU

We hope that someone makes a port of recent bitsandbytes to AMD/ROCM soon and we won't need this branch anymore (instead, you'd be able to run the main Petals version there).

[borzunov]

Not needed anymore, now the Petals main branch runs on AMD GPUs (even with NF4): https://github.com/bigscience-workshop/petals/wiki/Running-on-AMD-GPU

This is thanks to this bitsandbytes ROCM-compatible fork: https://github.com/arlo-phoenix/bitsandbytes-rocm-5.6