IMPORTANT NOTE: this is a fork of the original pytorch/csprng repo with some new useful features (in particular, reproducible CSPRNG).

This library has not received an audit, and you should think carefully before using it in a production system.

This fork of torchcsprng requires libsodium to work correctly; you should install it before installing this library.

Once you have libsodium installed, run:

pip install git+https://github.com/kernelmethod/csprng.git@main

Alternatively, add the following to your requirements.txt:

torchcsprng @ git+https://github.com/kernelmethod/csprng.git@main

This library exports one function, create_generator, which creates a new PyTorch Generator object with cryptographically secure pseudo-random number generation (CSPRNG).

>>> import torchcsprng as csprng
>>> gen = csprng.create_generator()

Note that the current implementation uses JIT-compiled extensions, so it may take a moment to compile the module the first time that you call it.

Once you've created your Generator instance, you can feed it to PyTorch's RNG functions, e.g.

>>> import torch
>>> torch.randn(3, 3, generator=gen)
tensor([[ 0.5859, -0.6674, -0.2723],
        [ 0.9075,  2.0364,  0.0331],
        [-0.2692, -1.1712, -1.5100]])

When you call create_generator without any arguments, it will automatically generate an secret key from the operating system's random stream. You can explicitly generate a key and pass it in to PyTorch in order to make RNG deterministic:

>>> import secrets
>>> key_bytes = secrets.token_bytes(32)
>>> key = torch.tensor([b for b in key_bytes], dtype=torch.uint8)
>>> gen = csprng.create_generator(key)