Kernels documentation
Build variants
Getting started
Using kernels
QuickstartUsing LayersLocking Kernel VersionsEnvironment VariablesMigrating from older versionsIntegrating kernels into a projectFAQ
Kernels API Reference
Kernels CLI Reference
Kernels CLIkernels benchmarkkernels versionskernels lockkernels downloadkernel-builder skillskernels verify-signature
Building kernels
Writing KernelsBuilding kernelsLocal DevelopmentIDE SetupKernel RequirementsSecurityWhy Nix?Metal NotesBuild VariantsBuilding kernels with agentsBuilder CLI Reference
Design
You are viewing main version, which requires installation from source. If you'd like
regular pip install, checkout the latest stable version (v0.15.2).
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
torch211-cpu-aarch64-darwintorch212-cpu-aarch64-darwin
Metal aarch64-darwin
torch211-metal-aarch64-darwintorch212-metal-aarch64-darwin
CPU aarch64-linux
torch211-cxx11-cpu-aarch64-linuxtorch212-cxx11-cpu-aarch64-linux
CUDA aarch64-linux
torch211-cxx11-cu126-aarch64-linuxtorch211-cxx11-cu128-aarch64-linuxtorch211-cxx11-cu130-aarch64-linuxtorch212-cxx11-cu126-aarch64-linuxtorch212-cxx11-cu130-aarch64-linuxtorch212-cxx11-cu132-aarch64-linux
CPU x86_64-linux
torch211-cxx11-cpu-x86_64-linuxtorch212-cxx11-cpu-x86_64-linux
CUDA x86_64-linux
torch211-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-linux
ROCm x86_64-linux
torch211-cxx11-rocm71-x86_64-linuxtorch211-cxx11-rocm72-x86_64-linuxtorch212-cxx11-rocm71-x86_64-linuxtorch212-cxx11-rocm72-x86_64-linux
XPU x86_64-linux
torch211-cxx11-xpu20253-x86_64-linuxtorch212-cxx11-xpu20253-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-xpu