To summarize my complaints:

all the stuff about installing the right versions of everything should be irrelevant; we should just make the system have working versions.
- For people who do want to install their own, we should provide a complete conda line that does it correctly, like pytorch does
tensorflow-gpu isn't necessary. Since (?? a while?) tensorflow and tensorflow-gpu both the GPU code, at least on Linux; judging from the filesizes and here the linux version always includes the GPU parts; plus the chart on https://www.tensorflow.org/install/source#gpu implies tensorflow 2.0 merged the GPU code into tensorflow itself. I think it can just tell people to pip install tensorflow.
The instructions need to specify a version: eg pip install tensorflow<=2.3 or pip install tensorflow~=1.15 because according to that same chart, the most recent ones are not compatible with CUDA 10.x
The note about CUDA being loaded to your .profile could be more accurate; it's actually installed via /etc/ld.so.conf.d/cuda-10-2.conf. This is effectively like setting LD_LIBRARY_PATH in your .profile, but it's not actually via your .profile. Anyway, I think we can just erase that whole note. But that's kind of irrelevant because:
CUDNN doesn't seem to be installed on rosenberg anymore? It doesn't show up as available with ldconfig -v -N 2>/dev/null | grep cudnn. If I drop it from my conda envs then tensorflow complains that it can't load it and doesn't detect any GPUs.
According to the chart on https://www.tensorflow.org/install/source#gpu, CUDNN 8 has to be paired with CUDA 11. But the SOP says rosenberg has CUDA 10 and CUDNN 8. So is that not true?
birelli has cuda 10.2 installed and nvidia 440, contradicting the docs.

Move the GPU documentation here #26

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions