Install

dmipy is a thin umbrella that pulls both engines:

pip install dmipy                 # dmipy-sim + dmipy-fit
pip install "dmipy[cuda12]"       # + JAX CUDA-12 GPU (JAX, jaxopt)
pip install "dmipy[cpu]"          # + JAX CPU

GPU ([cuda12])

Needs a working CUDA-12 stack, and the environment must export LD_LIBRARY_PATH so the dynamic loader resolves the CUDA libraries — see the dmipy-sim README for the exact export.

CPU-only / shared machine

Set JAX_PLATFORMS=cpu to force JAX onto CPU regardless of what is installed.

There is no importable dmipy package — import the engines directly:

import dmipy_sim          # forward Monte Carlo
import dmipy_fit          # analytical inverse

Upgrading from dmipy 1.x

dmipy 2.0 is a ground-up rewrite. The 1.x line (the original 2019 toolbox) was a single importable dmipy package; 2.x is a meta-package with no importable dmipy module. Old import dmipy code will not run under 2.x. To stay on the original toolbox, pin pip install "dmipy<2" — it remains available on PyPI.

Or install an engine on its own:

pip install dmipy-sim
pip install "dmipy-fit[jax]"