Triton (OpenAI)

Writing a custom GPU kernel without dropping into CUDA C++ is the whole point of Triton, a Python-embedded kernel language and compiler that started as Philippe Tillet's Harvard research, matured at OpenAI and is now a community project with its own developer conference. You write a triton.jit function that works on blocks of data, and the MLIR and LLVM based compiler handles memory access patterns, on-chip memory and scheduling before emitting PTX for NVIDIA or AMDGCN for AMD hardware. Fused attention, matmul and normalization kernels fit in a few dozen lines, and PyTorch's torch.compile emits Triton code for GPUs, which is why Linux builds of PyTorch pull the package in. Prebuilt wheels install with pip for CPython 3.10 through 3.14; Linux is the supported OS, with NVIDIA GPUs of compute capability 8.0 or newer and AMD GPUs on ROCm 6.2 or later, while CPU support is still in progress. Setting TRITON_INTERPRET=1 runs kernels in a CPU interpreter, handy for debugging and for the Triton Puzzles. The MIT license permits commercial use. Meant for ML systems engineers and researchers chasing speed that stock libraries cannot reach.