Kernels documentation
Build variants
Get started
Use kernels
QuickstartUse layersLock kernel versionsEnvironment variablesProjects using kernelsMigrate from older versionsKernels securityFAQ
Python API
kernels CLI
Overviewkernels cache lskernels infokernels describe-apikernels variantskernels versions (deprecated)kernels lockkernels downloadkernels benchmarkkernels verify-signature
Build kernels
Write kernelsBuild with NixDevelop locallySet up your IDESet up for Metal kernelsDevelop kernels with agentsSecure your kernelsGitHub Actions & HF Jobs
kernel-builder CLI
Advanced practices
Kernel specifications
Concepts & design
Resources
Build variants
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.
CPU aarch64-darwin
torch210-cpu-aarch64-darwintorch211-cpu-aarch64-darwintorch212-cpu-aarch64-darwintorch213-cpu-aarch64-darwintorch214-cpu-aarch64-darwintorch215-cpu-aarch64-darwin
Metal aarch64-darwin
torch210-metal-aarch64-darwintorch211-metal-aarch64-darwintorch212-metal-aarch64-darwintorch213-metal-aarch64-darwintorch214-metal-aarch64-darwintorch215-metal-aarch64-darwin
CPU aarch64-linux
torch210-cpu-aarch64-linuxtorch211-cpu-aarch64-linuxtorch212-cpu-aarch64-linuxtorch213-cpu-aarch64-linuxtorch214-cpu-aarch64-linuxtorch215-cpu-aarch64-linux
CUDA aarch64-linux
torch210-cu126-aarch64-linuxtorch210-cu128-aarch64-linuxtorch210-cu130-aarch64-linuxtorch211-cu126-aarch64-linuxtorch211-cu128-aarch64-linuxtorch211-cu130-aarch64-linuxtorch212-cu126-aarch64-linuxtorch212-cu130-aarch64-linuxtorch212-cu132-aarch64-linuxtorch213-cu126-aarch64-linuxtorch213-cu130-aarch64-linuxtorch213-cu132-aarch64-linuxtorch214-cu126-aarch64-linuxtorch214-cu130-aarch64-linuxtorch214-cu132-aarch64-linuxtorch215-cu130-aarch64-linuxtorch215-cu132-aarch64-linuxtorch215-cu134-aarch64-linux
CPU x86_64-linux
torch210-cpu-x86_64-linuxtorch211-cpu-x86_64-linuxtorch212-cpu-x86_64-linuxtorch213-cpu-x86_64-linuxtorch214-cpu-x86_64-linuxtorch215-cpu-x86_64-linux
CUDA x86_64-linux
torch210-cu126-x86_64-linuxtorch210-cu128-x86_64-linuxtorch210-cu130-x86_64-linuxtorch211-cu126-x86_64-linuxtorch211-cu128-x86_64-linuxtorch211-cu130-x86_64-linuxtorch212-cu126-x86_64-linuxtorch212-cu130-x86_64-linuxtorch212-cu132-x86_64-linuxtorch213-cu126-x86_64-linuxtorch213-cu130-x86_64-linuxtorch213-cu132-x86_64-linuxtorch214-cu126-x86_64-linuxtorch214-cu130-x86_64-linuxtorch214-cu132-x86_64-linuxtorch215-cu130-x86_64-linuxtorch215-cu132-x86_64-linuxtorch215-cu134-x86_64-linux
ROCm x86_64-linux
torch210-rocm70-x86_64-linuxtorch210-rocm71-x86_64-linuxtorch211-rocm71-x86_64-linuxtorch211-rocm72-x86_64-linuxtorch212-rocm71-x86_64-linuxtorch212-rocm72-x86_64-linuxtorch213-rocm71-x86_64-linuxtorch213-rocm72-x86_64-linuxtorch214-rocm714-x86_64-linuxtorch214-rocm72-x86_64-linuxtorch215-rocm100-x86_64-linuxtorch215-rocm714-x86_64-linux
TPU x86_64-linux
torch-tpu
XPU x86_64-linux
torch210-xpu20253-x86_64-linuxtorch211-xpu20253-x86_64-linuxtorch212-xpu20253-x86_64-linuxtorch213-xpu20260-x86_64-linuxtorch214-xpu20261-x86_64-linuxtorch215-xpu20261-x86_64-linux
Python-only kernels
Kernels 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-tputorch-xpu