I suggest you try mamba or micromamba, it’s way faster than vanilla conda
1 Like
The specific error you’re getting will be solved in Do not generate C code for BatchedDot when BLAS flags are missing by ricardoV94 · Pull Request #550 · pymc-devs/pytensor · GitHub
But the fact you hit it means your installation is suboptimal and in some models may result in much slower sampling (hence why we always suggest installing from conda instead of pip)
1 Like
This really helpful thanks - I didn’t know pip provides a slower sampling speed compared to conda. I’ll give mamba a shot and provide an update here.
Also, with your suggestion on updating to the latest pytensor, unfortunately, because of software restrictions, I have to stick with the version pytensor-rel-2.17.2 since installing a higher version requires microsoft visual build tools c++ 14.0 or greater which I’m not able to install.
That shouldn’t be a problem with conda/mamba right? If you have permissions to use it, it will take care of those sort of dependencies
When I run pip install <path-to-latest-pytensor-version.tar.gz file>, I get an error message (on my work computer) that mentions something along the lines of requiring Microsoft visual build tools C++ 14.0 or greater.
I’d have to try a fresh conda environment without any pip installs and then install pymc-dev packages via mamba. Will try it tomorrow.
1 Like
Hi @ricardoV94, I can confirm the code in Lkjcorr returning error - #17 by mcao as well as batching the dims in MvNormal works on my personal computer by pip installing the latest pytensor (version 2.18.3).
p.s. I haven’t tried conda/mamba in my work environment yet. I’ll mark this topic as solved since the issue is related to the work computer settings.
@mcao heads up we restricted the LKJCorr to the vector case in the latest PyMC release because it is not clear whether the logp for the batch case is actually doing the right thing