id	summary	reporter	owner	description	type	status	priority	milestone	component	version	resolution	keywords	cc	blockedby	blocking	notify_on_close	platform	project
17606	Install Boltz with the correct CUDA version	Tom Goddard	Tom Goddard	"Currently the Boltz tool installs the PyPi Boltz which on Windows includes torch with only CPU support.  If the Windows machine has Nvidia graphics it will run 25x faster using the gpu so I want to install pytorch with cuda support.  This is easy to do with a pip install from the pytorch repository as described at the pytorch install page

    https://pytorch.org/get-started/locally/

for example

    pip3 install torch --index-url https://download.pytorch.org/whl/cu126

The problem is there are 3 cuda versions 11.8, 12.6 and 12.8 to choose from.  I can determine what cuda version is supported by the nvidia driver on the computer in a few ways, either using the gputils pypi package (probably best) or looking at the graphics driver version reported by ChimeraX.  Once I know the system CUDA version, e.g. 12.4, I probably want to choose an older CUDA version of pytorch since CUDA is backward compatible.  This isn't completely clear.  On minsky in the vizvault on Linux the PyPi install of torch provides CUDA 12.6 but minsky has CUDA 12.4.  It still works fine in Boltz, probably because torch checks if the CUDA version has the features it needs and it does.  While pytorch cuda 11.8 might work it might not use some features that are in CUDA 12.4 that the 12.6 version will use.  Still probably safest to go with 11.8.  But a tricky problem is that 11.8, 12.6, and 12.8 are the currently supported pytorch versions, but 6 months from now they may have dropped 11.8 or changed the provided versions.  If ChimeraX hard-codes the available versions it may fail in the future if the pytorch site removes old versions, and it may not use new versions which exactly match the user's system CUDA.  So a problem is how I know programmatically which CUDA versions pytorch is distributing.

The same issues apply on Linux.  While the default PyPi torch linux package includes cuda 12.6 support, it probably would be more sensible to choose a compatible cuda.  If the Linux system appears to have no cuda I'm not sure I want a CPU torch because often linux nvidia drivers are not installed by mistake and later they get installed.  Still probably best in that case to install cpu torch since I have no way to know what cuda version the future installed nvidia driver will have.
"	enhancement	assigned	moderate		Structure Prediction								all	ChimeraX