WIP: Initial support for bnb 4bit on any nn.Parameter

[matthewdouglas]

What does this PR do?

This PR adds a new option to BitsAndBytesConfig called target_parameters with the same spirit as target_parameters in huggingface/peft#2638. The intent is to allow quantization of nn.Parameter that are not within a nn.Linear, e.g. those found commonly in certain MoE model implementations.

Requires bitsandbytes-foundation/bitsandbytes#1720 which is released in bitsandbytes v0.48.0.

Example usage with a Granite MoE:

model = GraniteMoeForCausalLM.from_pretrained(
    "ibm-granite/granite-3.1-3b-a800m-base",
    torch_dtype=torch.bfloat16,
    device_map="cuda:0",
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=False,
        target_parameters=["block_sparse_moe.input_linear.weight", "block_sparse_moe.output_linear.weight"],
        llm_int8_skip_modules=["lm_head", "block_sparse_moe.router"]
    ),
)

Memory Usage - BF16

Metric	Cur Usage	Peak Usage	Tot Alloc	Tot Freed
Allocated memory	6291 MiB	6292 MiB	12583 MiB	6292 MiB
Active memory	6291 MiB	6292 MiB	12583 MiB	6292 MiB
Requested memory	6291 MiB	6291 MiB	12583 MiB	6291 MiB

Memory Usage - Before PR

Metric	Cur Usage	Peak Usage	Tot Alloc	Tot Freed
Allocated memory	6019 MiB	6027 MiB	9935 MiB	3916 MiB
Active memory	6019 MiB	6027 MiB	9935 MiB	3916 MiB
Requested memory	6015 MiB	6024 MiB	9929 MiB	3913 MiB

Memory Usage - After PR

Metric	Cur Usage	Peak Usage	Tot Alloc	Tot Freed
Allocated memory	1894 MiB	2054 MiB	9424 MiB	7530 MiB
Active memory	1894 MiB	2054 MiB	9424 MiB	7530 MiB
Requested memory	1875 MiB	2035 MiB	9389 MiB	7513 MiB

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• [ x ] Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
  (See Slack discussion)
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@SunMarc @MekkCyber @BenjaminBossan

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Rocketknight1]

cc @MekkCyber

[SunMarc]

Nice it would be great to add some tests (inference / saving) with gptoss model !