Make cuda version check dynamic#202
Conversation
…ng the get_cuda_driver_script twice, as it's costly. We simply adapt the script to always return a 0 exit, and then do any handling of the case where EESSI_CUDA_DRIVER_VERSION is NOT set by the end in the calling Lmod hook
|
This now correctly sets Forcing it to e.g. With Just to test failures of the script, I modified the script and hardcoded an empty string for the driver version. Then I get the following, as expected: Setting that env var allowed me to load the module. |
|
bot: build repo:eessi.io-2023.06-software instance:eessi-bot-mc-aws for:arch=x86_64/amd/zen2 |
|
New job on instance
|
|
New job on instance
|
|
Ingested! |
properly deploy PR #202 (Make cuda version check dynamic)
| # This can be leveraged by the source_sh() feature of Lmod | ||
| # Because we want to source this without immediately raising an LmodError upon failure, this script | ||
| # is designed to ALWAYS return a 0 exit code | ||
| EESSI_CUDA_DRIVER_VERSION=$(nvidia-smi --query | grep -oP 'CUDA Version\s*:\s*\K[0-9.]+') || return 0 |
There was a problem hiding this comment.
This doesn't allow for the fact that someone can use compatibility libraries to improve the compute capability supported by the driver (and this can be done in user space)
There was a problem hiding this comment.
| EESSI_CUDA_DRIVER_VERSION=$(nvidia-smi --query | grep -oP 'CUDA Version\s*:\s*\K[0-9.]+') || return 0 | |
| EESSI_CUDA_DRIVER_VERSION=$(LD_LIBARARY_PATH=$EESSI_CUDA_COMPATIBILITY_LIBRARY_PATH nvidia-smi --query | grep -oP 'CUDA Version\s*:\s*\K[0-9.]+') || return 0 |
could give a way to do this
Fixes #189
First step to fixing #201
Unblocks #200 , which blocks EESSI/software-layer#1462 which blocks EESSI/software-layer#1453 ...
This PR can be modified by modifying the
create_lmodsitepackage.pywith something like: