tandem-torch

PyTorch tensors from Tandem8x32. CPU and CUDA fills write the stream of the specification bit for bit, fast on both.

API: the generator, dtypes, normals, exponentials, bounded draws and parallel use.
Design: the CUDA fills and the bounded, normal and exponential contracts.
Tests: what the suite checks.
Performance: M4 and A100 figures against torch.

This is not a torch.Generator. That class is final and its RNG hooks are internal to PyTorch, so no third-party generator can drive torch.rand. tandem_torch fills tensors through its own functions instead, which is also what makes the stream reproducible across languages and devices.

Install

git clone --recurse-submodules https://github.com/tandem-rng/tandem-torch
pip install torch                      # a CPU or CUDA build, see pytorch.org
pip install --no-build-isolation .

The reference sources sit in the external/tandem-c (1c75956) and external/tandem-cuda (e98daee) git submodules. Clone with git clone --recurse-submodules, or run git submodule update --init in an existing clone. GitHub's ZIP download omits submodules and does not build.

The build needs a C and C++ compiler. With nvcc on the path (or CUDA_HOME set) the CUDA fills are built too, otherwise CUDA tensors raise. Set TANDEM_TORCH_CUDA=0 or 1 to force the choice. For development, pixi install gives a CPU environment, pixi install -e cuda a Linux GPU environment with nvcc 12.8 from conda-forge and PyTorch's cu128 wheel, and pixi run test runs the tests.

AI assistance

This port was written with the help of large language models under human direction. The design and the specification are human work, as is much of the Julia implementation. The code is tested bit for bit against every vector of the specification and against long stream dumps from the Julia implementation, and every value must match. The output does not depend on who or what wrote the code.