| 31 Aug 2022 |
 tpw_rules | hexa: i don't 100% know how much in practice it matters or how to suggest it on github, but imo the python-packages.nix invocation of torch should have cudaPackages = cudaPackages_11_6; as that's what the wheels are built against and what they know they support | 18:13:19 |
| tpw_rules | Samuel Ainsworth: do you have an opinion on that? we should try to build with what upstream supports? | 18:18:38 |
| tpw_rules | * Samuel Ainsworth: do you have an opinion on that? imo we should try to build with what upstream supports? | 18:18:46 |
 Samuel Ainsworth | In reply to @me:linj.tech
could you provide a link? found it! https://cs.github.com/?q=repo%3Anixos%2Fnixpkgs+autoAddOpenGLRunpathHook | 18:48:05 |
| Samuel Ainsworth | In reply to @tpw_rules:matrix.org
Samuel Ainsworth: do you have an opinion on that? imo we should try to build with what upstream supports? the usual practice here is to build with the default cudaPackages or cudaPackages_11 if possible. if for whatever reason the package doesn't support the latest version, then we would override it with cudaPackages_11_6 etc | 18:50:16 |
| Samuel Ainsworth | that way users don't get stuck with a copy of every version of cuda in their /nix/store | 18:51:00 |
| tpw_rules | ok | 18:51:11 |
| Samuel Ainsworth | iirc the current cudaPackages is 11.6 atm anyhow | 18:51:18 |
| tpw_rules | yeah | 18:51:24 |
| tpw_rules | thought: say what you will about bazel, at least it knows how to load all my cores properly... | 18:51:51 |
| Samuel Ainsworth | yeah sometimes I really wish that nix would build things at a level more granular than whole packages... | 18:52:34 |
| Samuel Ainsworth | it can be so inefficient at times | 18:52:40 |
| tpw_rules | here i mean tensorflow vs pytorch's build system | 18:52:53 |
| Samuel Ainsworth | cough tensorflow mass rebuilds cough | 18:52:54 |
| Samuel Ainsworth | ahhh I see | 18:53:00 |
| tpw_rules | part of the problem is nvcc not doing all the different architectures in parallel | 18:55:00 |
| Samuel Ainsworth | mmm that's def annoying | 18:59:08 |
| tpw_rules | nvidia gpu accelerated compiler when | 19:23:16 |
| tpw_rules | 
Download image.png | 20:00:02 |
 hexa | tpw_rules: for both src and bin? | 21:33:12 |
| hexa | wait, the wheel doesn't consume it 😄 | 21:35:15 |
| tpw_rules | no i just tried src | 21:45:49 |
| tpw_rules | "consume"? | 21:46:00 |
| hexa | it does not have cudaPackages in its argument list | 21:48:18 |
| tpw_rules | it looks like it still includes cuda. let me see what happens... | 21:50:58 |
| hexa | the bin.nix doesn't mention cuda except in comments | 21:53:45 |
| tpw_rules | yeah ima try if i can get it to do cuda type things | 21:54:02 |
| tpw_rules | it should be much faster than building though | 21:54:11 |
| tpw_rules | yeah it does do cuda things | 21:56:01 |
| tpw_rules | and the wheel has all the cuda .so files | 21:56:24 |
