Installation
Simple Install
uv/pip (recommended)
Note
Cupy doesn't like nccl dependencies installed as wheels from pip. Importing
cupy the first time will present an import error with a python command that can
be used to manually install the nccl library, e.g. (for cuda 12.x - replace with
relevant cuda version)
For more up-to-date info, can follow the issue here.
Installing from source (development)
Clone the repo, then choose from the following:
uv
This will install the library in editable mode.Notes for development
- Installation with optional dependency specifiers:
[dev]- installs with development capabilities (e.g. docs, pytest, cov)[sigpy]- installs with sigpy to enable unit testing againstsp.ArrayToBlocksandsp.Interpolate- Careful doing this, as sigpy might not like certain package versions of things.
[all]- installs all optional dependencies!