[Release 0.3.0] Basic Readme and user-facing pathways

[dbogunowicz]

Feature Description

• Adds a README.md that, among other fundamental information, presents how to load/save single weight files as well as how to load/save models using compressed-tensors (note, there is no mention of AutoModels for now, we merely mimic how safetensors saves and loads state_dict for models). Those functions are implemented and tested as a part of this PR.
• Adds a examples/bitmask_compression.ipynb notebook, that explains in more detail how bitmask compression works
• Adds a make command to build a wheel
• Adds three user-facing functions load_compressed, save_compressed, save_compressed_model, that mimic the UI introduced by safetensors.
• Adds an exhaustive set of tests to validate the aforementioned functionalities

User Facing Functions

load_compressed(
    compressed_tensors: Union[str, Path], # path to a single file or directory with multiple .safetensor files 
    compression_config: CompressionConfig = None,
    device: Optional[str] = "cpu",
) -> Generator[Tuple[str, Tensor], None, None]: # returns a generator for memory efficiency
...

def save_compressed(
    tensors: Dict[str, Tensor],
    save_path: Union[str, Path],
    compression_format: Optional[CompressionFormat] = None,
) -> None: # saves to disk
...

def save_compressed_model(
    model: torch.nn.Module,
    filename: str,
    compression_format: Optional[CompressionFormat] = None,
    force_contiguous: bool = True,
) -> None: # saves to disk

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB