Hello.
Im very glad for this new release of CNTK2.7
But actually, pip install cntk-gpu require to install also manually a lot of dll dependencies, where the wheel of CNTK2.6 was all in one with full dependencies.
Is a complete wheel file will be released or do i install all dependencies cudatoolkit/cudnn/mkl/opencv manually?
Thx
How do we install the dependencies?
I solved it by copying them manually from the binary distribution...
I solved it by copying them manually from the binary distribution...
I tried the same things.
I downloaded this release : https://cntk.ai/dlwg-2.7.html (1Gb ...)
And this wheel file : https://cntk.ai/PythonWheel/GPU/cntk_gpu-2.7-cp36-cp36m-win_amd64.whl (94 mb ...)
I opened the .whl file with 7Zip, then i copied the missing 10 dependencies :
cublas64_100.dll
cudart64_100.dll
cudnn64_7.dll
curand64_100.dll
cusparse64_100.dll
libiomp5md.dll
mkldnn.dll
mklml.dll
nvml.dll
opencv_world310.dll
from the folder "\cntk\cntk" of the release to the folder "\cntk\libs\" of the opened wheel archive.
Finally, i can install the new wheel file with :
pip install cntk_gpu-2.7-cp36-cp36m-win_amd64.whl
And everything works. This wheel can now be reused for differents conda env etc... (its easier than rebuilding setup.py but the sha256 integrity of the wheel be broken)
I can close the issue.
@xgirones workaround is good.
I will double check whether we can repack CNTK 2.7 to include dependencies really soon
@xgirones workaround is good.
I will double check whether we can repack CNTK 2.7 to include dependencies really soon
thx a lot <3
We just repacked CNTK 2.7. Enjoy!
Most helpful comment
We just repacked CNTK 2.7. Enjoy!