-
Notifications
You must be signed in to change notification settings - Fork 66
New issue
Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? # to your account
Windows support #43
Comments
Hello, there is a development branch 4f1e6ba which solves the fnctl issue, but this isn't the only one. Pure C++ or numpy part of keops works fine for now. Note that ninja, cmake, a working visual studio 2019 and llvm/clang compiler are required. |
On hold : due to quarantine, I lost my access to a gpu for windows for testing the port. |
Hi All, You may have seen my mention for this issue above. PyKeOps is a dependancy in a recently released Gaussian Process Regression package, and I was just wondering if installing the above development branch will provide me with a Windows compatible version? Apologies if you haven't managed to get round to it yet @fradav. |
Windows support for this package would be much appreciated! Just posting to affirm that there is interest. |
Due to various reasons, I'm withdrawing from the project from now on, starting a brand new project with similar goals. Regards, |
@fradav is Windows WIP available somewhere? |
What is the status of this issue? Is there any plan for adding Windows support for pykeops? |
Thanks for your interest in the library! Best regards, |
Hi @jeanfeydy, thanks for continuing to work on Windows Integration. I have tried the steps mentioned in #281 but I encountered an error. Would appreciate any help regarding this, thanks! |
Hi @An-u-rag, can you provide the cuda, pytorch version that you tested? |
Thanks for the reply @initialneil. Pytorch version is 1.13.1 and Cuda version is 11.7. The test_keopscore.py file tries to import "get_compiler_library" function from the keops package located in miniconda3 site-packages directory. When I navigate to the given location's config.py file and look for it manually, I see that there is no implementation for that function in the config.py file. However, when I access the local keopscore directory in the cloned keops local repo, I can find a similar config.py file which does seem to have the implementation for the same function. I am not sure why this is the case. Would it be fine for me to copy this config.py file to the miniconda's site-package version of it or am I doing something wrong? |
@initialneil Yep that was the issue. Fixed it with pip uninstall. Thank you for all the help! I'm not sure if this is a general concern or is it because of your changes to keops for windows but I am using a LazyTensor instantiation with Vi and Vj in my project and I seem to be getting these warnings from keops, is there a way I can suppress them? |
@An-u-rag Right cause this version is not officially tested and merged yet. These are normal logs for verbose purposes. |
@initialneil I tried working with it but I keep getting errors. Now, I can't even get the initial import to work due to a circular import error. I am not sure what is happening. I might just wait for the official release with windows support. Thanks for the help anyway! |
Hi @jeanfeydy, I get the same error described initially by @lelayf concerning the missing module named fcntl and required for KeOps to work on Windows. Many thanks for your time in advance ! |
Hi @Antoine-insa , Thanks for your interest in our library!
In any case, you should be able to try the branch by following the instructions here. Please feel free to let us know about your experience with it. |
Many thanks @jeanfeydy , I will try the instructions you've pinpointed. |
@jeanfeydy Hello, do you know if there are any updates to have it work on Windows? Unfortunately, I get the following error [KeOps] Generating code for formula Sum_Reduction((Var(0,3,0)-Var(1,3,1))|(Var(0,3,0)-Var(1,3,1)),1) ... OK [KeOps] error: cuMemcpyDtoH(out, (CUdeviceptr) out_d, sizeof(TYPE) * sizeout) failed with error CUDA_ERROR_INVALID_VALUE I do not know how to solve that because it’s beyond my knowledge of CUDA and C++ programming. I using PyTorch 2.5.1 and CUDA 12.4. Many thanks. Really looking forward to try this out! |
a litlle more search shows that
fcntl
is only supported on Mac/Linux. It seems it is only used in one place to lock a file, which could be done in a more portable way withportalocker
?The text was updated successfully, but these errors were encountered: