// the take
Warp is NVIDIA's Python framework for writing GPU kernels as plain Python functions that get JIT-compiled to CPU or CUDA. It's aimed at people doing physics simulation, robotics, or differentiable ML pipelines who want kernel-level performance without hand-writing CUDA C++.
Kernels are differentiable out of the box and plug into PyTorch/JAX/Paddle autograd, so you can backprop through a physics step instead of treating simulation as a black box. The tile-based GPU programming primitives (gemm, cholesky, fft, nbody) go beyond particle-sim toy examples into actual linear algebra building blocks. It's backed by real engineering process — per-PR changelog fragments via towncrier, CI with codecov, and steady commit activity rather than a research-paper dump.
GPU acceleration is CUDA-only — macOS wheels run CPU-only with no Metal path, so Apple Silicon users get none of the speedup that's the whole point of the library. Hardware/toolkit versioning is a headache: default PyPI wheels target CUDA Toolkit 13.4 and require a Turing (sm_75)+ GPU, so anyone on CUDA 12 or older hardware has to hunt down a separate +cu12 wheel or build from source. The license is more tangled than the Apache 2.0 badge implies — prebuilt wheels statically link NVIDIA's libmathdx under its own separate license terms, which is worth checking before you assume this is cleanly Apache-licensed end to end.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →