torch-check
torch-check inspects your Python and NVIDIA environment and recommends a compatible wheel from the official PyTorch index. Recommendations support CPython on Linux x86_64 with glibc, with a CPU fallback when no reviewed CUDA wheel qualifies.
Install
Install to ~/.local/bin using a POSIX shell. Pass --version or --install-dir after sh -s -- to customize the installation.
|
Or install from crates.io:
GitHub Releases provides checksummed binaries for Linux x86_64 (tested on Ubuntu 22.04), Linux Arm64, macOS, and Windows x64. Use the Windows archive from PowerShell or Command Prompt. Linux Arm64, macOS, and Windows builds support inspection.
Usage
Use --python /path/to/python to target an interpreter, --gpu 0,2 to select physical GPU indices from nvidia-smi, and --format json for structured output. Run torch-check --help for all options.
Choose pip, uv, or uv-add with --installer. Generated commands target the inspected Python environment. verify imports the installed torch and runs checks on the selected GPUs. Use --redact before sharing reports. See SECURITY.md for security details.
Compatibility
Wheel tags, glibc, GPU architecture, and the NVIDIA driver determine compatibility. The CUDA version displayed by nvidia-smi is informational. Official wheels can run without a local CUDA Toolkit.
direct-compatible meets the normal driver minimum. minor-compatible relies on CUDA minor-version compatibility, so PTX JIT or newer driver features can still fail. unverified indicates missing evidence, and verified requires a successful runtime check. Use candidates --unverified to include unverified builds.
Wheel metadata is cached for 24 hours. Use --refresh to fetch it again or --offline to use a complete local snapshot.
Primary sources (generated from the metadata files; do not edit this block manually):
Reviewed compatibility sources — latest registered PyTorch series: 2.15. A registered configuration is not proof of wheel availability or GPU architecture coverage; those are checked separately.
- NVIDIA CUDA minor-version compatibility
- NVIDIA CUDA 13.2.0 release driver table
- PyTorch release matrix
- PyTorch CUDA architecture matrix (ref v2.15.0)
- PyTorch CUDA architecture matrix (ref v2.14.0)
- PyTorch CUDA architecture matrix (ref efe07ed21d89)
- Previous PyTorch versions
- PyTorch 2.15 release build matrix
- PyTorch 2.14 release build matrix
- PyTorch 2.11.0 Linux wheel build configuration
- PyTorch 2.10.0 Linux wheel build configuration
- PyTorch 2.9.1 Linux wheel build configuration
- PyTorch 2.9.0 Linux wheel build configuration
- PyTorch 2.8.0 Linux wheel build configuration
- PyTorch 2.7.1 Linux wheel build configuration
- PyTorch 2.7.0 Linux wheel build configuration
- PyTorch 2.6.0 Linux wheel build configuration
- TorchVision 0.29 version metadata
- PyTorch issue #190355
Observed upstream sources — latest observed PyTorch series: 2.15. Observations are not automatically approved compatibility rules, including when a document also appears in the reviewed list.
- NVIDIA CUDA minor-version compatibility
- NVIDIA CUDA 13.2.0 release driver table
- PyTorch release matrix
- PyTorch CUDA architecture matrix (ref v2.15.0)
- PyTorch CUDA architecture matrix (ref v2.14.0)
- PyTorch CUDA architecture matrix (ref efe07ed21d89)
- PyTorch CUDA architecture matrix
- PyTorch 2.11.0 Linux wheel build configuration
- PyTorch 2.10.0 Linux wheel build configuration
- PyTorch 2.9.1 Linux wheel build configuration
- PyTorch 2.9.0 Linux wheel build configuration
- PyTorch 2.8.0 Linux wheel build configuration
- PyTorch 2.7.1 Linux wheel build configuration
- PyTorch 2.7.0 Linux wheel build configuration
- PyTorch 2.6.0 Linux wheel build configuration
Wheel availability is queried separately from the PyTorch official wheel index.
License
This project is distributed under either the MIT License or the Apache License, Version 2.0, at your option. See LICENSE for details.