Install Torch

Install Torch

REQUIRED BACKGROUND: Read and follow $use-pixi before changing a Pixi manifest.

Install all three PyTorch distributions only after the caller supplies the target platform and one of: CPU, MPS, or a concrete CUDA wheel channel such as cu130.

Required input

Collect these values before editing:

• selected Pixi manifest and environment
• Pixi platform: linux-64, win-64, or osx-arm64
• interpreter family, Python major/minor version, and expected Python/ABI tags; the inspector supports standard GIL-enabled release CPython only, derives cpNN from major/minor, and accepts that Python tag with cpNN, abi3, or none ABI
• explicit backend: CPU, MPS, or a concrete CUDA channel

If the backend is not explicit, stop and invoke $setup-torch; do not inspect GPU hardware here.

This template uses standard GIL-enabled release CPython. Refuse automated inspector validation for PyPy, free-threaded tags such as cp313t, debug ABI tags such as cp313d, or other interpreter variants. Do not pass 3.13 and silently treat a variant as the standard cp313 target.

Resolve a compatible trio

Before choosing versions, read PyTorch compatibility sources and follow its official source ordering.

1. Read the current PyTorch previous versions page. Treat one published row/command as the authoritative torch / torchvision / torchaudio release group.
2. Prefer the newest complete release group supporting the requested Python, platform, and backend. A newer torch release without a matching torchaudio release is not a complete trio.
3. Map CPU to https://download.pytorch.org/whl/cpu. For macOS/MPS, use the default PyPI simple index unless the official command says otherwise. Map CUDA channel cuXYZ to https://download.pytorch.org/whl/cuXYZ.
4. Inspect each package index directly. A reproducible first check is:

curl -fsSL https://download.pytorch.org/whl/cu130/torch/
curl -fsSL https://download.pytorch.org/whl/cu130/torchvision/
curl -fsSL https://download.pytorch.org/whl/cu130/torchaudio/

5. From the repository root, run the inspector with every required input. For example:

python .agents/skills/install-torch/scripts/inspect-wheel-index.py \
  --index-url https://download.pytorch.org/whl/cu130 \
  --python-version 3.13 \
  --platform linux-64 \
  --torch-version 2.11.0 \
  --torchvision-version 0.26.0 \
  --torchaudio-version 2.11.0 \
  --json

For macOS/MPS, pass --index-url https://pypi.org/simple. Exit 0 means a complete trio exists. Exit 1 means a wheel is missing or the candidates have no channel-consistent trio. Exit 2 means invalid input, a network or inspection failure, or multiple compatible local-version channels are ambiguous and require an explicit +local version. For exits 0 and 1, inspect the JSON compatible, missing, matches, selected_channel, and channel_issue fields; exit 2 reports the inspection error on stderr. 6. Stop on a missing package. Do not mix release rows, change Python silently, or omit torchaudio.

Change the Pixi manifest

Pin the verified versions under the selected platform's pypi-dependencies. In pixi.toml, a CUDA target looks like:

[target.linux-64.pypi-dependencies]
torch = { version = "==2.11.0", index = "https://download.pytorch.org/whl/cu130" }
torchvision = { version = "==0.26.0", index = "https://download.pytorch.org/whl/cu130" }
torchaudio = { version = "==2.11.0", index = "https://download.pytorch.org/whl/cu130" }

The root [target.<platform>.pypi-dependencies] table belongs to the default feature. It affects the selected environment only when that environment includes the default feature.

If the selected environment sets no-default-feature = true, or Torch should remain optional, put the pins under a named feature already owned by that environment:

[feature.torch.target.linux-64.pypi-dependencies]
torch = { version = "==2.11.0", index = "https://download.pytorch.org/whl/cu130" }
torchvision = { version = "==0.26.0", index = "https://download.pytorch.org/whl/cu130" }
torchaudio = { version = "==2.11.0", index = "https://download.pytorch.org/whl/cu130" }

Here, torch is an example owning feature name that must already be present in the selected environment. Do not silently change environment membership: choose an owning feature already present in the environment, or ask before changing the environment definition. For a Pixi manifest embedded in pyproject.toml, prefix the whole table path with tool.pixi, producing [tool.pixi.target.linux-64.pypi-dependencies] for the default-feature example or [tool.pixi.feature.torch.target.linux-64.pypi-dependencies] for the named-feature example. Adapt the concrete platform and feature names to the selected environment. For macOS/MPS, omit a custom index when the official command uses PyPI. Change only the requested target; do not add a CUDA toolkit unless the project independently requires local CUDA compilation.

Lock and verify

Run:

pixi lock --manifest-path pixi.toml
pixi lock --manifest-path pixi.toml --check
pixi run --manifest-path pixi.toml --environment default --platform linux-64 --locked python -c "import torch, torchvision, torchaudio; print(torch.__version__, torchvision.__version__, torchaudio.__version__)"

Replace pixi.toml, default, and linux-64 with the selected manifest, environment, and platform. pixi lock resolves all declared platforms and environments in the workspace. Run the import command only on a host or CI runner compatible with the selected platform; a non-host pixi run --platform invocation does not prove runtime usability. --locked prevents verification from silently rewriting a stale lock file.

Verify the installed versions belong to the selected release group. Report the manifest diff, owning feature and environment, index, Python/platform tags, lock result, and import versions. Do not report device availability here; $setup-torch owns device verification.