MPS progress

[tcdent]

Got Metal -> pyobjc++ -> Python loading. quantize_mps doesn't segfault. Start of quantize_nf4 in the public API.

Builds with: $ cmake -DCOMPUTE_BACKEND=mps -S . && make

import ctypes
import torch
from bitsandbytes import lib
from bitsandbytes.functional import get_ptr, get_4bit_type

lib_path = "bitsandbytes/libbitsandbytes_mps.dylib"
lib = ctypes.CDLL(lib_path)

blocksize, n, quant_type = 64, 1024, "nf4"
A = torch.rand(n).float()
code = get_4bit_type(quant_type, device=A.device)
absmax = torch.zeros(n // blocksize).float()
out = torch.zeros((n + 1) // 2).byte()

lib.quantize_mps(
    get_ptr(code),
    get_ptr(A),
    get_ptr(absmax),
    get_ptr(out),
    ctypes.c_int32(blocksize),
    ctypes.c_int(n)
)
print(out[:10])

tensor([ 51, 51, 255, 255, 51, 51, 51, 51, 51, 51], dtype=torch.uint8)

[Titus-von-Koeller]

Hey @tcdent, really cool, looks nice! Happy to see that you're taking the initiative 🤗

However, any specific reason why you merged main? We're keeping things separate for the time being, so this messes up the diff and we also don't want everything from main in multi-backend-refactor; especially not mixed in with a PR, makes it hard to review.

Happy to give this a look soon. Better to revert the merge from main though, if you can.

[tcdent]

Sorry, sloppy on my part. This work is a tangent of some tests I was trying to do, which needed both branches to be up-to-date.

Is a revert OK? I can also re-open a fresh PR.

[Titus-von-Koeller]

@tcdent Could you please update us on the status of this PR. Is this work still foreseen, are the benefits of what's there already self-contained and ready to be merged?

Sorry for these questions, we have a huge task load right now, in need for triage, and I would like to get your opinion on this before looking deeper into the matter.

Thanks so much for your initiative on this PR in the first place!