Pycuda install failing

Install, Configure and Update
1

I don’t know how to report to the build team, but they may be interested in seeing this for possible updates to the build process.

Tried another clean install on ldap 15.1, --standalone installation.

I’d have sworn last attempt installed both client and server, but I had to run ./install.sh in both master and worker directories, not sure what I did differently.

Issue as before was pycuda, though I have I believe, resolved installation issues that I had been previously unaware of.

This is the result of # pip install pycude, an immediately previous uninstall told me pycuda was not installed.

I believe I need pycuda to properly connect the worker code to the GPUs, please let me know if I’m wrong or there is another way.

This was run as the cryosparc user immediately after I found that I wasn’t able to run cryosparc2.
I was in the worker directory at the time of execution.

I am including the complete install log, I apologize for the length.

Nope, deleted all of that, found this very interesting link.

However as with the cuda 10.1 (currently highest cuda installed on system)
we are finding #include crt/host_config.h fails.

Nope, but using cuda 10.1. Failing to include files using CUDA_INC_DIR, performing a copy of .h files to /usr/include as long as they don’t interfere with existing files, one needed a new subdirectory /usr/include/crt

I recognize this work-around as evil, but need to get our Cryosparc users running again.

Got down the the linker, can’t find curand, I assume curand.so, which I can not find for cuda 10.1.

2

Did another completely fresh install, this time against cuda 8.0 (which we had been using for Cryosparc 2.9). Fine install, up until the attempted connect of the worker with the master - which is still necessary on a -standalone build?

bin/cryosparcw connect failed to find the pycuda.driver, but it reminded me that last week were were able to manually invoke python3 and import pycuda.driver Wondering if this isn’t a simple correctable path issue and not actually an issue with pycuda install.

pycudea uninstall and reinstall has been miserable, but I wonder if the attempt is even necessary.

Is anyone else installing/updating cryosparc v2.9 to v2.11 on Leap 15, with GPUs?

thanks,
Brian

3

Noted that anaconda has python 2.7 modules, moved /usr/bin/python softlink to python 3.6m executable and deleted and did a completely fresh install. No change to the python that anaconda installed.

install.sh script did run the worker script this time, it hasn’t been doing that lately, and we did catch very specific errors for the python build, linker failed on “collect2” with lib64/crti.o file not recognized, failed format not recognized.

Looks like its pulling from cryosparc2_master/deps/anaconda/computer_compat/ld directory, which is essentially empty.

thanks in advance,
Brian

4

Hey team, I am having a similar issue. While I’m able to compile and install worker on nodes using Cuda-9.2, when I try to install/use cuda 10.1 I cannot (pycuda wheel fails). While for the most part it is ok, unfortunately we have some RTX 2080 GPU, which do not work with cuda versions below 10.
Is there anyone who can give me an idea of what I need to be doing?
Thanks