CMake fails for CUDA 11

[alchernyavskiy]

Hi,

Running pip install or cmake from source produces the following error for CUDA 11:

nvcc fatal : Unsupported gpu architecture 'compute_30' failed

Does anyone know how to fix it?

[neubig]

Try removing the parts of this line that refer to compute_30:
https://github.com/clab/dynet/blob/master/dynet/CMakeLists.txt#L265

We may have to update this file to cover more recent architectures, please tell me if this works and I'll do so if so.

[alchernyavskiy]

Try removing the parts of this line that refer to compute_30:
https://github.com/clab/dynet/blob/master/dynet/CMakeLists.txt#L265

We may have to update this file to cover more recent architectures, please tell me if this works and I'll do so if so.

I managed to build it for cuda 11 without any errors. However, during python import it prints
ImportError: libcublas.so.10: cannot open shared object file: No such file or directory

I used the following command for cmake:
cmake .. -DEIGEN3_INCLUDE_DIR=../../eigen -DPYTHON='which python' -DBACKEND=cuda -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-11.1 -DCUDA_ARCH=61

[heeh]

Same error. Removing compute_30 does not help.

nvcc fatal : Unsupported gpu architecture 'compute_30'
CMake Error at dynet_generated_gpu-ops.cu.o.Release.cmake:215 (message):
Error generating
/tmp/pip-install-9ldty3ty/dynet/build/py3.7-64bit/dynet/CMakeFiles/dynet.dir//./dynet_generated_gpu-ops.cu.o

[heeh]

Was able to change it to compute_60, which corresponds to NVIDIA tesla P100
Also, the Eigen repository is broken; had to download manually.
wget https://github.com/clab/dynet/releases/download/2.1/eigen-b2e267dc99d4.zip

[tranhoangkhuongvn]

I have this exact error when trying to compile on GPU RTX3090, any help would be appreciated!
I tried the above 2 suggestions by either removing compute_30 or replace it with compute_60 but it does not fix the error.

Thanks