cufile-python

A basic Python wrapper for the NVidia cuFile API

Installation

pip install cufile-python

Usage

Basic usage with the CuFile context manager:

from cufile import CuFile
import torch
import ctypes

# allocate an empty buffer in VRAM
t = torch.empty((1024, 1024, 16), dtype=torch.float32, device="cuda")
with CuFile("test.bin", "r") as f:
    f.read(ctypes.c_void_p(t.data_ptr()), t.nbytes)

Alternatively one could import cufile.bindings and use it as if calling libcufile from C++.

Development

To set up the development environment:

1. Clone the repository:

git clone https://github.com/yanok/cufile-python.git
cd cufile-python

2. Install development dependencies:

pip install -r requirements.txt

3. Run tests:

python -m pytest tests/

License

This project is licensed under the MIT License - see the LICENSE file for details.