Agustinus Kristiadi

New library by PyTorch: * https://pytorch.org/blog/torchtune-fine-tune-llms/ * https://github.com/pytorch/torchtune First thoughts: * The model is compatible. Notice that their LoRA uses nn.Linear https://pytorch.org/torchtune/stable/tutorials/lora_finetune.html * The dataset should also be compatible as...

Make `*Laplace` respect the `dtype` of the base model

2

Closes #246

Outputs from `mkdocs`: ``` WARNING - griffe: laplace/baselaplace.py:1971: Parameter 'num_data' does not appear in the function signature WARNING - griffe: laplace/baselaplace.py:1971: Parameter 'diagonal_kernel' does not appear in the function signature...

Functional samples

1

Closes #241. **Use case:** computation of an arbitrary expectation w.r.t. $p(f(x) \mid x, \mathcal{D})$.

### Simplify typehinting Typehinting in the docstring is redundant. **Before** ```python def f(x: Tensor, n: int = 5) -> Tensor: """ Some paragraph. Parameters ---------- x: Tensor Some words. n:...

Once #242 is merged, we have a nice doc and thus can simplify the currently huge README.md.

Strict static typing with MyPy => better API for the user Example problems when no typechecking is performed: #233

Currently, the way we inherit `BaseLaplace` is kinda weird, resulting in:

Currently, we require the user to do this themselves: ```python class MyGPT2(nn.Module): def __init__(self, tokenizer: PreTrainedTokenizer) -> None: super().__init__() config = GPT2Config.from_pretrained("gpt2") config.pad_token_id = tokenizer.pad_token_id config.num_labels = 2 self.hf_model =...

Support for multidim outputs

2

Closes #252 Thanks to Kazuki for updating `asdfghjkl` on PyPI, we can now handle 3D tensor outputs (e.g. in LLMs) with GLM predictive.