Apex: Given no hashes to check 137 links for project 'pip': discarding no candidates

[keloemma]

hello, I am trying to install apex following your recommendation and I get this error, I try ot uninstall and reinstall but I laways get this error :

Given no hashes to check 137 links for project 'pip': discarding no candidates
when looking on my entired stacktrace I found this other error :

I am not really an expert so do you know I can solve it.

I am trying to install in a virtaul env : with version of torch (1.3.1) and cuda 10.1

[maxjiang93]

Same issue as this: #690

[maxjiang93]

What worked for me was to switch to cuda 10.1 from cuda 10.2 that I was using:

export CUDA_HOME=/usr/local/cuda-10.1

[keloemma]

How can I switch to cuda 10.1 to 10.2 should I reinstall torch ?

[keloemma]

What worked for me was to switch to cuda 10.1 from cuda 10.2 that I was using:

export CUDA_HOME=/usr/local/cuda-10.1

How can I switch to cuda 10.1 to 10.2 should I reinstall torch ?

[GuardedAirplane]

So I found the answer on another thread. It seems the best way of installing is to execute the following: python setup.py install --cpp_ext --cuda_ext.
The caveat is that you need to have a system cuda version that matches the pytorch version you downloaded.

[songsuoyuan]

So I found the answer on another thread. It seems the best way of installing is to execute the following: python setup.py install --cpp_ext --cuda_ext.
The caveat is that you need to have a system cuda version that matches the pytorch version you downloaded.

You definitely need gcc version >= 5. required

[thammegowda]

• gcc >= 4.9 and gcc <= 8 (as informed by warning in the logs); Check gcc --version
• set export CUDA_HOME=/usr/local/cuda-<version> variable; use to the proper cuda for which pytorch was built
• To know the cuda version to which your pytorch was built: python -c 'import torch; print(torch.version.cuda)'

P.S. I tried this on PPC64le CPU arch (torch 1.3, cuda 10.2, GCC 7 .. ), and I am yet to test if this actually works! But the pip list | grep apex showed it as installed, so I consider the build and installation as a success.

[rubick1896]

I am still getting this error on torch 1.5.1. My apex worked fine when my torch version was 1.3.1. I updated torch version recently and apex stopped working. So I try to re-install apex and gets this error.

nvcc version: 10.1.105
install pytorch via: conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

I'm sure my CUDA_HOME was set correctly.

[GuardedAirplane]

What happens when you type nvcc --version?