Closed jxtps closed 9 months ago
Switching to the cxx11 ABI version solved the issue, but the recommended usage style with cuda-platform-redist
should probably be fixed regardless?
nvrtc-builtins
is apparently missing from the list of libraries to preload for pytorch. I'll add it to PR #1466 .
In the meantime, you can manually load class org.bytedeco.cuda.global.nvrtc
from your code.
But I'm not sure it will be enough because it's used by libnvrtc.so
and cuda libraries do not have their RPATH set to $ORIGIN
. @saudet, do you have a mean to patch the cuda libraries and add the RPATH before building the native jar ?
Commit pushed to preload nvrt-plugins
.
Forget what I said about RPATH. preload actually loads the library, not only extracts it in the cache, so setting the RPATH shouldn't be needed.
I'm running into some problems deploying the (as of this writing latest) version of https://github.com/bytedeco/javacpp-presets/tree/master/pytorch
2.1.2-1.5.10
.If I try to use libtorch (https://download.pytorch.org/libtorch/cu121/libtorch-shared-with-deps-2.1.2%2Bcu121.zip), then I get:
Not sure where
_ZNK3c106Device3strB5cxx11Ev
would be defined? Hmm... maybe I should try the cxx11 ABI version? (i.e. https://download.pytorch.org/libtorch/cu121/libtorch-cxx11-abi-shared-with-deps-2.1.2%2Bcu121.zip)If I switch to using
then I get:
when I try to actually load a pytorch model.
Looking in
/mnt/lib/cache/org.bytedeco.cuda-12.3-8.9-1.5.10-linux-x86_64-redist.jar/org/bytedeco/cuda/linux-x86_64$
I can see:i.e.
libnvrtc.so
is there, but nolibnvrtc-builtins.so.12.3
.???