-
Notifications
You must be signed in to change notification settings - Fork 22.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Bizarre "no kernel image" error for pytorch built from source #32759
Comments
det is implemented in MAGMA, so this might be related to how you compiled against magma. Could you please run the environment collection script |
Very insightful. I remember seeing some documentation on the MAGMA gitlab page allowing specification of CUDA compute capabilities, I'm using the pre-built ones from the pytorch conda channel. I will try and build without MAGMA.
|
It does seem to be a MAGMA problem. Closing now. |
Modified the patch code to add support for compute capability 3.5
And rebuilt pytorch again, worked like a charm |
I followed this document - https://github.com/pytorch/pytorch/#from-source and build it.
|
Hi @aluo-x - #32759 (comment) |
Our cluster has retired GPUs from the K80 generation, I believe we target Nvidia 1080Ti~3090 now - so this is no longer a problem for me, and I no longer build Pytorch/MAGMA from source. But installing magma from conda will not work for you, you will need to clone it from here then specify the correct arch here. |
🐛 Bug
A workstation with Nvidia k40c (compute capability 3.5), Anaconda 2019.10, python 3.7.6, Ubuntu 19.10, GCC 8.3.0, driver 440.48.02.
CUDA SDK version 10.1.243, and cudnn 7.6.5.
Pytorch 1.4.0 is built from source using instructions according to those available on github, with Magma 101 installed.
export TORCH_CUDA_ARCH_LIST="3.5"
is set.Compilation succeeds without fatal errors.
The following works:
The following fails:
It is super odd that adding a single dimension causes the determinant to fail. I can reproduce the issue using Pytorch 1.3.1 and 1.4.0. The same operation works on the CPU. More information can be provided if needed.
Edit:
The vast majority of operations work. I can run complex 100M plus parameter networks with instance/layer/spectral/batch norm, as well as applying gradient penalties to a discriminator.
Edit:
cc @ezyang @ngimel
The text was updated successfully, but these errors were encountered: