Skip to content

v0.1.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Jul 12:20
· 50 commits to main since this release
v0.1.0
3908ef2

Initial production release of torch-check, a CLI tool that inspects a Linux/Python/NVIDIA environment and recommends an installable official PyTorch wheel.

Highlights

  • Detects OS, architecture, kernel, glibc, CPython ABI, Python executable, NVIDIA GPUs, compute capability, NVIDIA driver, driver-reported CUDA, and local CUDA Toolkit.
  • Resolves against official PyTorch wheel indexes instead of assuming a wheel exists.
  • Recommends CUDA wheels from reviewed PyTorch release metadata, with CPU fallback only when no reviewed CUDA candidate passes static checks.
  • Correctly treats nvidia-smi's reported CUDA version as informational, not as a hard upper bound for PyTorch wheel selection.
  • Supports NVIDIA CUDA minor-version compatibility rules and warns when PTX JIT or newer-driver features may still require a driver upgrade.
  • Includes reviewed GPU architecture evidence for PyTorch 2.6 through 2.13, including Blackwell sm120 coverage where supported.
  • Generates install commands for pip, uv pip, and uv add.
  • Defaults install commands to the active Python environment; --python pins detection, verification, and install command generation to a specific interpreter.
  • Provides compact human output, structured JSON output, candidate inspection, explanation mode, and runtime verification mode.
  • Uses atomic metadata cache writes and offline/refresh controls.
  • Ships without telemetry and never executes installation commands.

Supported production target

The production recommendation target is Linux x86_64 with glibc, CPython, NVIDIA GPUs, and official PyTorch pip wheels.

Other platforms can run the binary and inspect the host, but PyTorch wheel recommendation is intentionally limited where the compatibility model is not yet supported.

Prebuilt binaries

This release includes archives for:

  • x86_64-unknown-linux-musl
  • aarch64-unknown-linux-musl
  • x86_64-apple-darwin
  • aarch64-apple-darwin
  • x86_64-pc-windows-msvc

Each archive contains the binary, shell completions, a man page, project licenses, security policy, and third-party license notices. SHA256SUMS is included for archive verification.

Install

From crates.io:

cargo install torch-check --locked

Or download a prebuilt archive from this release.

Example

torch-check
torch-check --installer uv
torch-check candidates --unverified
torch-check explain torch==2.13.0 --cuda cu126
torch-check verify

Notes

This is the first public release. Compatibility rules are intentionally conservative: uncertain CUDA candidates are shown for review, but the default install command is reserved for candidates backed by reviewed metadata and static checks.

Assets 8
Loading