Wheel Variants for PyTorch Release 2.10 Published

atalman · February 5, 2026, 4:34pm

What are Wheel Variants?

Wheel variants are a mechanism for publishing platform-dependent Python wheels and selecting the most suitable package variant for a given platform.
This approach removes the need for manual selection of local identifiers in PyTorch packaging and greatly enhances the user experience when installing PyTorch.
For technical details and ongoing discussion, please see PEP-817 and associated discourse thread

What is currently supported?

The following variants are supported for PyTorch 2.10.0:

CUDA 12.6, 12.8 and 13.0: Linux x86, Linux aarch64, Windows x86
ROCm 7.0 and 7.1
XPU: Windows x86 and Linux x86
CPU only: Linux, Linux aarch64, MacOS (including M1/M2), Windows (CPU only)

How to test it?

You can start using Wheel Variants today by installing the PyTorch 2.10.0 release with the variant-enabled uv installer, which will automatically select the best PyTorch build for your machine’s hardware.

Using Linux/MacOS:

curl -LsSf https://astral.sh/uv/install.sh | INSTALLER_DOWNLOAD_URL=https://wheelnext.astral.sh/v0.0.3 sh
uv pip install torch torchvision

Using Windows:

powershell -ExecutionPolicy Bypass -c "$env:INSTALLER_DOWNLOAD_URL='https://wheelnext.astral.sh/v0.0.3'; irm https://astral.sh/uv/install.ps1 | iex"
uv pip install torch torchvision

If you have any questions or feedback, please feel free to reach out.

Cheers,
Team PyTorch

dvrogozh · February 12, 2026, 9:41pm

@atalman, thank you for posting this! I wonder 2 things:

Is CI/CD now covering the variant wheels at least for release procedure to make sure things are working on the real hardware? I believe you were working on the automation…
Are releases with variant wheels now automated? Will we see 2.10.1, 2.11.0, etc. coming out per Pytorch release schedule or there is still manual work which needs to happen for variant wheels releases?

atalman · February 16, 2026, 3:40pm

HI @dvrogozh
The release of the variants is automated and done by this workflow:

github.com/pytorch/test-infra

.github/workflows/release-wheel-variants.yml

main

name: Release Wheel Variants

on:
  pull_request:
    paths:
      - .github/workflows/release-wheel-variants.yml
      - release/pypi/upload_variant_to_staging.sh
  workflow_dispatch:
    inputs:
      pytorch_version:
        description: 'PyTorch version to promote (e.g., 2.10.0)'
        required: true
        type: string
      pytorch_release:
        description: 'PyTorch release series for config files (e.g., 2.10)'
        required: true
        type: string
        default: '2.10'
      package_name:
        description: 'Package name to promote'

This file has been truncated. show original

The validations of these wheels are done using following workflow:

github.com/pytorch/test-infra

.github/workflows/validate-binaries.yml

main

name: Validate binaries

# A reusable workflow that triggers a set of jobs that perform a smoke test / validation of pytorch binaries.
# Optionally restricts validation to the specified OS and channel.
# For the details about parameter values, see:
#   pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
# For an example of the `workflow_call` usage see:
#   https://github.com/pytorch/builder/pull/1144
on:
  workflow_call:
    inputs:
      os:
        description: "Operating system to generate for (linux, windows, macos, macos-arm64)"
        required: true
        type: string
      channel:
        description: "Channel to use (nightly, test, release, all)"
        required: true
        type: string
      torchonly:

This file has been truncated. show original

We do relay on the override flags for all GPU types, here:

github.com/pytorch/test-infra

.github/scripts/validate_binaries.sh

main


      
          get_test_suffix() {
              if [[ ${TORCH_ONLY:-} == 'true' ]]; then
                  echo "--package torchonly"
              else
                  echo ""
              fi
          }
          
          # Configure environment variables for wheel variants based on GPU type
          configure_wheel_variant_env() {
              case ${MATRIX_GPU_ARCH_VERSION:-} in
                  12.6*)
                      export NV_VARIANT_PROVIDER_FORCE_CUDA_DRIVER_VERSION='12.6'
                      export NV_VARIANT_PROVIDER_FORCE_SM_ARCH='6.0'
                      ;;
                  12.8*)
                      export NV_VARIANT_PROVIDER_FORCE_CUDA_DRIVER_VERSION='12.8'
                      export NV_VARIANT_PROVIDER_FORCE_SM_ARCH='9.0'
                      ;;
                  13.0*)
                      export NV_VARIANT_PROVIDER_FORCE_CUDA_DRIVER_VERSION='13.0'

Functionality wise these wheels are exactly the same as our regular wheels released to pypi and download.pytorch.org and these are covered by PyTorch CI system.

Topic		Replies	Views
Discover a simpler way to install PyTorch tailored to your hardware: Wheel variants for Release 2.8 are now available for testing on the PyTorch test channel release/packaging	0	810	August 2, 2025
PyTorch 2.10 Final RC Available Release Announcements	2	742	January 19, 2026
PyTorch 2.10 RC1 produced for pytorch, torchvision & torchaudio Release Announcements	0	810	December 12, 2025
PyTorch Release 2.1.0 - Final RC is available Release Announcements	0	2570	September 13, 2023
PyTorch 2.11 RC1 produced for pytorch & torchvision Release Announcements	0	363	February 19, 2026

Wheel Variants for PyTorch Release 2.10 Published

What are Wheel Variants?

What is currently supported?

How to test it?

Related topics