pytorch
libtorch C++ windows: The specified module could not be found. mkl_vml_def.1.dll
🐛 Describe the bug
Downloaded libtorch C++ windows CPU version:
https://download.pytorch.org/libtorch/cpu/libtorch-win-shared-with-deps-2.3.0%2Bcpu.zip
While attempting to run a program built against libtorch I get:
INTEL MKL ERROR: The specified module could not be found. mkl_vml_def.1.dll.
Intel MKL FATAL ERROR: cannot load mkl_vml_def.1.dll.
It seems the distribution is missing necessary dependencies.
Versions
libtorch version 2.3.0
downloaded via https://download.pytorch.org/libtorch/cpu/libtorch-win-shared-with-deps-2.3.0%2Bcpu.zip
cc @peterjc123 @mszhanyi @skyline75489 @nbcsm @vladimir-aubrecht @iremyux @Blackhex @cristianPanaite @jbschlosser
I was able to solve it myself, so I will provide more information which may help you deal with the issue.
The distribution is missing MKL runtime dlls. It contains mkl_core.1.dll and mkl_intel_thread.1.dll but I couldn't find any information on the website or github about which version of MKL it was built with. By looking at the aforementioned dll metadata, it looks like the file versions is 2021.4.1.0. The easiest way for me was to grab corresponding nuget package of mkl runtime here : https://www.nuget.org/packages/intelmkl.redist.win-x64/2021.4.0.640. I just downloaded the file, unzipped it and put the runtime dlls from runtimes/win-x64/native along with the rest of them. It's not exactly the same version, but it works fine now and with good performance (though I noticed the avx512 dlls were not used).
This issue is also present in the Cuda 12.1 Windows distribution of 2.3.0. @bsekura suggestion solved the issue for Cuda 12.1 as well.
Downloaded via https://download.pytorch.org/libtorch/cu121/libtorch-win-shared-with-deps-2.3.0%2Bcu121.zip
Thank you so much @bsekura, I was several hours into debugging. This is also an issue in the CUDA 11.8 distribution of 2.3.0 on Windows, both the release and the debug version. I am noticing that the performance of my program is quite a bit slower than it was on previous versions, although I don't really know if this is the cause. I will probably try downgrading to a different version of LibTorch for now.
Using an older worked for me. I'm on cpu though but the corresponding cuda version might work too:
https://download.pytorch.org/libtorch/cpu/libtorch-win-shared-with-deps-2.2.2%2Bcpu.zip https://download.pytorch.org/libtorch/cu121/libtorch-win-shared-with-deps-2.3.0%2Bcu121.zip
Just replace the version number for 2.2.2 or older. Now I can matmul two tensors yay! haha
@bsekura this issue should be fixed in 2.4.0. Can you check with that version and let me know if you have issues?
I am closing this issue due to a lack of activity and since this should be fixed from 2.4.0. Please reopen a new issue if this problem persists.