cudaPackages.cudnn: migrate to redist cuda, fix missing zlib#168748
cudaPackages.cudnn: migrate to redist cuda, fix missing zlib#168748FRidh merged 7 commits intoNixOS:masterfrom
Conversation
|
Since I've removed |
This comment was marked as outdated.
This comment was marked as outdated.
5818564 to
a1feea7
Compare
|
Ok, autoPatchelf has removed |
There was a problem hiding this comment.
I'd move it to args, but I'd have to split 8.3.2.44 into two instances - pre and post cuda 11.4 - and adjust the selector expression smh... I'll keep it this way, I guess
|
@SuperSandro2000 @FRidh would you have any objections now? |
There was a problem hiding this comment.
May seem redundant, but the idea is to repeat a pattern used elsewhere anyway, e.g. in pytorch. The pattern being: that packages use cuda as a "toolchain" with chosen hosts and targets, and we decide in the derivation which modules to include in the toolchain. Runfile being a maybe-non-redist throw-all-in toolchain.
Higher-level goal: I want people like Sandro (maintainers but out of narrow cuda context) not to get confused when just skimming through changes
|
@FRidh @SuperSandro2000 I'd like to either get new change requests or a merge, because right now pytorch and cudnn are still broken (runtime errors) on master1 Footnotes
|
|
warnings are not allowed and need to be removed. Other than that, it looks good to me. |
aaf5abb to
f8c947b
Compare
f8c947b to
4f9ca6b
Compare
|
I removed the warnings, rebased and squashed the branch to remove the noise, fixed an error with It seems the full EDIT: actually why, I'll just give katago the old cudatoolkit in |
instead of custom find ... -exec ldd | grep routine mark libcudnn_cnn_infer.so as needed for libcudnn.so on cudnn>=8.0.5 - a hint for autoPatchelf, as an alternative to manually adding $ORIGIN
as a more common way to use addOpenGLRunpath and autoPatchelf with cudaPackages
...since cudnn is part of the cuda package set - introduces the scary useCudatoolkitRunfile function argument to discourage usage of the runfile-based cudatoolkit - instead of the rather hidden useRedist term in let ... in - repeats cudatoolkit_root pattern after cuda_joined in pytorch &c (the "toolchain view") - redist packages are marked optional to support cuda<11.4 where the attributes for redist packages do not exist
Co-authored-by: Sandro <sandro.jaeckel@gmail.com>
4f9ca6b to
3e3e1e2
Compare
|
|
@FRidh Ready |
samuela
left a comment
There was a problem hiding this comment.
changes LGTM. @FRidh @SuperSandro2000 are we good to merge?
Description of changes
cudatoolkit, except for temporarily keeping it inpropagatedBuildInputscheckPhasethat comparesRunpathagainstDT_NEEDEDvialddautoPatchelfHook?RunpathThe mentioned missing dependency leads to runtime errors when evaluating jit-traced models in cuda-enabled pytorch
CC @NixOS/cuda-maintainers
Things done
sandbox = trueset innix.conf? (See Nix manual)nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage./result/bin/)nixos/doc/manual/md-to-db.shto update generated release notes