GpuArrayException (b'GPU is too old for CUDA version')

[piyushrungta25]

Hi I installed libgpuarray and pygpu along with theano following instruction from docs.
When I run

import pygpu
pygpu.test()

I get the following error

pygpu.gpuarray.GpuArrayException: b'GPU is too old for CUDA version'

Same error when importing theano with device = cuda0 in .theanorc.

My cuda installation is working fine(verified by making and running the deviceQuery from cuda samples).

Video card - Nvidia Geforce 820m
nvidia driver - Driver Version: 390.59
CUDA Driver Version / Runtime Version - 9.1 / 8.0
OS - Manjaro linux

Some search showed that 820m supports CUDA 8.0. And my gpu is clearly not too old for the CUDA since the samples are running fine. Searching for the exception dosent turn up any relevant results so any help here would be appreciated.

[nouiz]

Maybe our error message isn't clear, but you have this error: *CUDA Driver Version / Runtime Version* - 9.1 / 8.0 The driver installed and running should be the same for libgpuarray to work well. A reboot could help you fix this problem.

…

On Sun, Jun 17, 2018 at 8:51 AM Piyush rungta ***@***.***> wrote: Hi I installed libgpuarray and pygpu along with theano following instruction from docs <http://deeplearning.net/software/libgpuarray/installation.html>. When I run import pygpu pygpu.test() I get the following error pygpu.gpuarray.GpuArrayException: b'GPU is too old for CUDA version' Same error when importing theano with device = cuda0 in .theanorc. My cuda installation is working fine(verified by making and running the deviceQuery from cuda samples). *Video card* - Nvidia Geforce 820m *nvidia driver* - Driver Version: 390.59 *CUDA Driver Version / Runtime Version* - 9.1 / 8.0 *OS* - Manjaro linux Some search showed that 820m supports CUDA 8.0. And my gpu is clearly not too old for the CUDA since the samples are running fine. Searching for the exception dosent turn up any relevant results so any help here would be appreciated. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#573>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALC-0zoB5cC4jko3CpD0vKJwrfupclgks5t9lDagaJpZM4Uq0ra> .

[nuqz]

@nouiz thanks for your tip about driver versions. Finally got Theano working after installing drivers from cuda toolkit run package.

[susanwl]

Hi,

I have absolutely the same issue with the same same gpu.

My setup is nvidia-drivers : 390.87 and cuda : 8.0.61. And deviceQuery
gives:

CUDA Driver Version / Runtime Version 9.1 / 8.0

I believe that the current check for device compatibility is not fully correct,
since:

1. CUDA Driver driver version apparently indicates not the currently supported
   cuda version, but instead the maximum cuda version that is currently
   supported:
   see

   That is, CUDA Driver Version 9.1 supports all cuda versions from 1 to 9.1.

2. I have manually disabled compatibility check in gpuarray_buffer_cuda.c
   and compiled libgpuarray without it. Without that check pygpu.test()
   completes almost without errors:

Ran 7253 tests in 258.780s

FAILED (SKIP=7, errors=5)

Among these 5 errors, 4 are due to missing libnccl.so and one:

File "/usr/lib64/python3.6/site-packages/pygpu/reduction.py", line 263, in call
raise ValueError("Array too big to be reduced along the

Correspondingly, I believe that the compatibility check gives false positive
for this setup, and probably should be rewritten to check for cuda runtime
version instead of driver version.

[dsimba]

Hi,

I have absolutely the same issue with the same same gpu.

My setup is nvidia-drivers : 390.87 and cuda : 8.0.61. And deviceQuery
gives:

CUDA Driver Version / Runtime Version 9.1 / 8.0

I believe that the current check for device compatibility is not fully correct,
since:

1. CUDA Driver driver version apparently indicates not the currently supported
   cuda version, but instead the maximum cuda version that is currently
   supported:
   [see](https://devtalk.nvidia.com/default/topic/1032152/cuda-driver-version-9-0-cuda-runtime-version-8-0/)
   That is, CUDA Driver Version 9.1 supports all cuda versions from 1 to 9.1.

2. I have manually disabled compatibility check in `gpuarray_buffer_cuda.c`
   and compiled `libgpuarray` without it. Without that check `pygpu.test()`
   completes almost without errors:

hi, would you mind sharing the exact details of the change you did to disable this version check?

thanks

[susanwl]

Hi,

I am not sure how detailed I should be, but I have just commented out that check in src/gpuarray_buffer_cuda.c. Here is the exact patch I have used against v0.7.6
patch