When PyTorch supports CUDA 10.1?

From NVIDIA,It seems that the CUDA 10.1’s performance has a significant improvement over CUDA 10, When pytorch supports CUDA 10.1 ?



I am pretty sure you can compile from source for CUDA 10.1 with no problem. Let us know here if you have any issue with it.
Building binaries for all cuda versions is a bit complex as they are so many of them, that is why there is only a few of them supported as binaries.

1 Like

Got it! Thank you very much!

I actually have the same question and wanted to clear something out. The documentation mentions that we have to run the following command to install dependencies conda install -c pytorch magma-cuda92

Should we run this with magma-cuda10, since magma-cuda101 is not available?


I have verified that one cannot build PyTorch with CUDA 10.1, after a few rigorous attempts. The main reason is that MAGMA does not support 10.1 yet.


Will you please report back when you managed to build PyTorch with CUDA 10.1? Thanks in advance.

Does not #17874 show that it is possible to install PyTorch with CUDA 10.1?

It really does’t at all. Even if the building/installation succeeds, the actual usage will get you errors as well. I forgot exactly which error (I attempted this a couple of months back), but it’s unsolvable until official magma-cuda101 becomes available.


Thank you very much this is very helpful.

Thanks for telling us! Should really be set somewhere clear that CUDA 10.1 won’t work, tried for a long time to make it work :sweat_smile:

Does anyone know when CUDA 10.1 will be supported with pip install torch commands? Thanks.

Hi, I also wait for cuda 10.1 support, because I couldn’t manage to install cuda 10.0 on Ubuntu 19. I bought a RTX 2070 just for deep learning and up to now the only use it has, is the LED bar on its side :frowning:

It seems, that magma 10.1 is available now:

Mangma 10.1

@jottbe Is there a reason you need CUDA 10.1 specifically? We are using PyTorch with CUDA 10.0 and a RTX 2080 and it is running just fine (both pytorch 1.0 as well as 1.1).

Mi Martin, there is no particular reason, only that I can’t install CUDA 10.0. It is not in the repository of Ubuntu 19 unfortunately and I already spent a lot of time trying to install everything.
I also would like to get tensorflow running with cuda. I know it currently also supports only 10.0, but if I now compile something myself, it will probably make some other manual changes necesary and getting everything clean again once it is officially supported, so I can go with the standard, is difficult sometimes.

The very strange thing is, that it seems to have run a few days ago for pytorch (not for tf), not sure what happend inbetween. Very frustrating.

Any update about when will CUDA 10.1 be supported with PyTorch? Is the support scheduled in the coming month(s)? Thanks.

Think it already is now. Despite being built on 10.0, it runs happily with 10.1. Here’s my env, which largely follows from the nvidia/cuda:10.1-devel-ubuntu18.04 image.

PyTorch version: 1.2.0
Is debug build: No
CUDA used to build PyTorch: 10.0.130

OS: Ubuntu 18.04.3 LTS
GCC version: (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0
CMake version: Could not collect

Python version: 3.7
Is CUDA available: Yes
CUDA runtime version: 10.1.243
GPU models and configuration: GPU 0: GeForce RTX 2080 Ti
Nvidia driver version: 430.40
cuDNN version: Could not collect

Versions of relevant libraries:
[pip] numpy==1.16.4
[pip] torch==1.2.0
[pip] torchfile==0.1.0
[conda] blas                      1.0                         mkl  
[conda] mkl                       2019.4                      243  
[conda] mkl-service               2.0.2            py37h7b6447c_0  
[conda] mkl_fft                   1.0.14           py37ha843d7b_0  
[conda] mkl_random                1.0.2            py37hd81dba3_0  
[conda] pytorch                   1.2.0           py3.7_cuda10.0.130_cudnn7.6.2_0    pytorch
[conda] torchfile                 0.1.0                    pypi_0    pypi

I believe it is now possible because cuda-magma101 has been released!