Publish the documentation to readthedocs.org and link to it from the README.

[hodgestar]

What will make this tricky is to build the documentation without cupy being present, but probably we can make a plan.

[MrRobot2211]

We could probably build the documentation with cupy installed but with it not being importable/runnable see this P.S.
#19 (comment)

[leofang]

You can build CuPy in RTD mode by setting CUPY_NO_CUDA=1. This way it's importable.

[leofang]

You can build CuPy in RTD mode by setting CUPY_NO_CUDA=1. This way it's importable.

Sorry, to be precise it's READTHEDOCS=True pip install cupy, which sets the C macro CUPY_NO_CUDA during the build.

[MrRobot2211]

Thank you @leofang. I think that when using a conda installation without a GPU it is defaulting to that installation. Could it be? See #19 (comment). When running the linter there are no import failures.

[MrRobot2211]

Thank you @leofang. I think that when using a conda installation without a GPU it is defaulting to that installation. Could it be? See #19 (comment). When running the linter there are no import failures.

On a second thought falke8 can't show the import failure, so disregard this.

[leofang]

I think it's two separate questions. The conda-forge packages are built for GPUs with a particular version of CUDA. But, your question reminded me that recent CuPy versions (probably staring v9) have strived to make import cupy not fail even if the CUDA installation is crippled or missing. If this is all we need (not running actual GPU code, just importing for RTD) I guess it's probably fine to just install CuPy as you'd do under normal circumstances.