Installation

Two dependencies. No CUDA. No vendor lock-in.

Vision needs Python 3.10 or newer and a working Vulkan, Metal or DX12 driver. It has two dependencies, numpy and wgpu; nothing from CUDA or ROCm.

From the repository

pip install ".[proof]"        # or: uv pip install ".[proof]"

The proof extra adds cryptography, used to sign and time-stamp receipts. Without it, everything else works.

The distribution is called vision-gpu, because the name vision is already taken on PyPI. The import is import vision and the command is vision.

Check the devices

vision devices

[0] NVIDIA GeForce RTX 4050 Laptop GPU — DiscreteGPU via Vulkan (driver: NVIDIA 595.99.02)
[1] Intel(R) Graphics (RPL-P) — IntegratedGPU via Vulkan (driver: Intel open-source Mesa driver Mesa 25.2.8-0ubuntu0.24.04.2)
[2] llvmpipe (LLVM 20.1.2, 256 bits) — CPU via Vulkan (driver: llvmpipe Mesa 25.2.8-0ubuntu0.24.04.2 (LLVM 20.1.2))
[numpy] reference on the CPU, no GPU (the same arithmetic, bit by bit)

Without a GPU, llvmpipe (the Mesa software Vulkan driver on Linux) and the numpy reference give the same results, only slower.

Check this hardware

vision conformance

The conformance kit runs 60 cases on every device and compares each one with reference signatures: the .vis language, the tensor library, the integer (MX) path and a SHA-256 on the GPU. Same command, same answer on NVIDIA, Intel, llvmpipe, numpy and Apple Silicon. See Conformance kit.

For development

uv venv .venv && uv pip install --python .venv/bin/python -e ".[proof]"
.venv/bin/python -m unittest discover -s tests -v