A kernel can be compliant for a specific compute framework (e.g. CUDA) or architecture (e.g. x86_64). For compliance with a compute framework and architecture combination, all the build variants listed below must be available. This list will be updated as new PyTorch versions are released.
torch211-cpu-aarch64-darwintorch212-cpu-aarch64-darwintorch211-metal-aarch64-darwintorch212-metal-aarch64-darwintorch211-cxx11-cpu-aarch64-linuxtorch212-cxx11-cpu-aarch64-linuxtorch211-cxx11-cu126-aarch64-linuxtorch211-cxx11-cu128-aarch64-linuxtorch211-cxx11-cu130-aarch64-linuxtorch212-cxx11-cu126-aarch64-linuxtorch212-cxx11-cu130-aarch64-linuxtorch212-cxx11-cu132-aarch64-linuxtorch211-cxx11-cpu-x86_64-linuxtorch212-cxx11-cpu-x86_64-linuxtorch211-cxx11-cu126-x86_64-linuxtorch211-cxx11-cu128-x86_64-linuxtorch211-cxx11-cu130-x86_64-linuxtorch212-cxx11-cu126-x86_64-linuxtorch212-cxx11-cu130-x86_64-linuxtorch212-cxx11-cu132-x86_64-linuxtorch211-cxx11-rocm71-x86_64-linuxtorch211-cxx11-rocm72-x86_64-linuxtorch212-cxx11-rocm71-x86_64-linuxtorch212-cxx11-rocm72-x86_64-linuxtorch211-cxx11-xpu20253-x86_64-linuxtorch212-cxx11-xpu20253-x86_64-linuxKernels that are in pure Python (e.g. Triton kernels) only need to provide one or more of the following variants:
torch-cputorch-cudatorch-metaltorch-rocmtorch-xpu