CUDA 10.0 Linux Build

[CMobley7]

I see there is a CUDA 10 Window's build. Is there no CUDA 10 Linux's build?

[zeyiwen]

The default pip install should work fine. If you want to build a wheel file for your specific environment, please refer to build-pyhon-wheel-file link for detailed instructions.

[sensharma]

@zeyiwen Hi, I had the same question as OP. The pip install does not work with CUDA 10.0 as you suggest. It gives me the error: libcusparse.so.9.0: cannot open shared object file: No such file or directory which indicates that it is built with CUDA 9.0. Is there a python 3.7 CUDA 10.0 linux build available or would we have to build it ourselves?

[zeyiwen]

Currently, you need to build one by yourself. We will build a wheel file for CUDA 10.0. Please stay tuned.

[zeyiwen]

Hi @CMobley7 @sensharma Would you be able to build a wheel file of CUDA 10 and share it with us? We can upload it to the repository.

[sensharma]

Hi @CMobley7 @sensharma Would you be able to build a wheel file of CUDA 10 and share it with us? We can upload it to the repository.

@zeyiwen I'm very new to this but I've attempted to build a wheel with py37 and cuda10 on linux (ubuntu18.04). All 34 test cases on https://github.com/Xtra-Computing/thundersvm/blob/master/docs/how-to.md and the 2 examples on https://github.com/Xtra-Computing/thundersvm/tree/master/python seem to be running fine.

But I don't know how to check if it is using the gpu (I think it uses it by default?) or if it is correctly using cuda10. Is there something in the API that can be used to check this?

The test output is attached if you want to have a look. Can't attach the .whl here. Let me know how to share it so you can inspect.

tsvm_test_output.txt

[zeyiwen]

It appears that your program only uses CPUs. Please clone a new ThunderSVM and follow the instruction here to build the wheel file.

After building the wheel file, please replace the thundersvm-0.3.3-py3-none-linux_amd64.whl file under python/dist/ directory of ThunderSVM.

You can open another terminal and execute watch nvidia-smi to monitor if your program uses GPUs.

[sensharma]

Those were the instructions I followed in the first place. Tried it again - same result. Doesn't seem to be using the GPU. Have attached the make output - there is a warning in it that may tell you what might be going wrong. Any flag I need to use during make that is not mentioned in the instructions?

tsvm_make_output.txt

[CMobley7]

@sensharma, if you're still having problems tomorrow let me know and I'll make the wheel file.

[sensharma]

@CMobley7 I am still stuck in the same situation. Would be great if you could have a go at your end.

[ghost]

@CMobley7 could you provide your wheel file, please. I have the same problem as @sensharma I get the error: libcusparse.so.9.0: cannot open shared object file: No such file or directory

and would it be possible for me to use python 3.6? or do your wheel file will only support python 3.7? Thanks

[1113571287]

Who can provide the wheel file for CUDA10 under the Linux system ?

[mjanson]

Here is a wheel file I generated per the directions above.

wheel_file.zip

I have verified

$ ldd libthundersvm.so | grep cuda libcusparse.so.10 => /usr/local/cuda/lib64/libcusparse.so.10 (0x00007f5a10105000)

@zeyiwen can you please incorporate this into a new release?

[zeyiwen]

Sure. If it works, we will add it in the new release.

[GregSilverman]

Any word on the new release? I'm in a restricted VM behind a proxy server with no easy way to get files sin, other than pip install. Thanks!

[mcamplaniROVCO]

Hi Everybody,

I compiled correctly the libraries for cuda 10.20. But when using the libraries the error I got is this one. Do you have any suggestion to solve this?

FATAL [default] Check failed: [error == cudaSuccess] CUDA driver version is insufficient for CUDA runtime version

[zhDai]

Hi Everybody,

I compiled correctly the libraries for cuda 10.20. But when using the libraries the error I got is this one. Do you have any suggestion to solve this?

FATAL [default] Check failed: [error == cudaSuccess] CUDA driver version is insufficient for CUDA runtime version

Have you fixed it？I also cuda 10.2

[icobx]

Hi everyone,

I am trying to build thundersvm on cuda 11.0, but i always get:

/home/jovyan/dp/thundersvm-master/src/thundersvm/kernel/kernelmatrix_kernel.cu:5:10: fatal error: cusparse.h: No such file or directory #include <cusparse.h> ^~~~~~~~~~~~ compilation terminated.

I have tried multiple posted fixes and solutions but nothing seems to be working. Do you have any ideas?