CUDA

CUDA runs supported Float32 operations on an NVIDIA GPU without changing the model’s tensor types or graph. The linked guide chapters give the full provider and trust account; the material below collects the commands used to build and test the native path.

Build and Run

Build with CUDA support:

lake -R -K cuda=true build

Run a small CUDA example:

lake -R -K cuda=true exe torchlean mlp --device cuda --steps 100

Run a broader CUDA regression pass:

scripts/checks/example_regression.sh --cuda

Run the CUDA sanitizer suite when changing native kernels:

scripts/checks/cuda_sanitize_tests.sh --all-tools

What CUDA Covers

The CUDA path is used for supported Float32 tensor operations: elementwise arithmetic, reductions, matmul/cuBLAS paths, convolution/pooling kernels, shape/view operations, attention kernels, FFT/FNO support where enabled, and model examples that choose --device cuda.

Determinism

Some CUDA reduction and backward paths use floating-point accumulation. Because Float32 addition is not associative, atomic accumulation can be schedule-dependent. TorchLean also provides an opt-in deterministic reductions mode for the covered reduction, gather/scatter, and pooling-backward paths:

let _ := Runtime.Autograd.Cuda.Buffer.setDeterministicReductionsChecked true

or:

TORCHLEAN_CUDA_DETERMINISTIC_REDUCTIONS=1 lake -R -K cuda=true exe torchlean mlp --device cuda

The setting covers the reduction, gather/scatter, and pooling-backward paths named in the GPU chapter, which also explains what the CUDA tests establish. For provider selection, VJP ownership, and assurance policies, read Inside the Backend Planner.