cudaPackages: bump default cudaPackages_11_7 -> cudaPackages_11_8 #224927
Conversation
|
Looks good... as long as nixpkgs-review is happy, we're good to go! |
|
I didn't think through the scope of nixpkgs-review for bumping CUDA. 11.8 built fine, but now it's downloading 80G worth of stuff to my machine. This may not happen tonight, but I guess we'll see. |
|
Yeah, running nixpkgs-review is the real hurdle with these version bumps |
|
Does it have a check-point mechanism? If I stop and restart tomorrow, will it pick up where it left off? |
|
Not really, but builds and downloads that finish will stay cached in your nix store (as long as you don't garbage collect them!). |
|
Oh good, thanks for confirming. I'll have to wrap it up tomorrow probably. |
|
Result of 55 packages failed to build:
44 packages built:
|
Failed derivations
|
ofborg
bot
added
8.has: clean-up
8.has: package (new)
|
Thanks @SomeoneSerge. The failure list looks pretty bad. In the case of xgboost (which I see in the fail list), I had intended to use the cudaPackages argument to fix the package to something that worked, and I already know the next version mostly works with 11.8. What do you suggest we do now? |
|
@nviets hmmm, I don't see most of the build logs, got to retrieve them somehow now xD |
|
Result of 55 packages failed to build:
44 packages built:
|
|
Oh!
It's #224986 as well 🤦🏻 |
|
Will #224986 directly fix cudatoolkit? If it works, do you want to bump to 11_8 or go straight to 12? |
|
@nviets that should fix the patchelf issues in I'd say we keep waiting for pytorch and such before we do cudaPackages_12 |
|
Given PyTorch's sensitivity to the CUDA version, would it make sense to make it configurable there? |
It is, one can always |
|
Result of 13 packages failed to build:
86 packages built:
|
|
Looking much better! How can i see the logs from your build? Sorry I couldn't help more after the initial commit, but my computer doesn't have the horsepower to build all of the downstream stuff. |
|
They were supposed to be automatically published, but smth went wrong and I'm omw to get lost in a Finnish "forest"/national park so it's hard for me to check up on it xD Maybe somebody coild rebuild individual packages, cc @samuela @nixos/cuda-maintainers |
|
Fwiw jaxlib had been failing last time I checked, idk if anyone fixed it yet |
|
The other failures aren't new either |
|
Sounds like an awesome trip! Thanks for running another nixpkgs-review. If this PR isn't breaking anything new, I think it's ready for review. Happy to recieve more feedback. |
|
tried running nixpkgs-review, ran into Mic92/nixpkgs-review#328 |
nviets
removed request for
layus,
jtojnar,
teto,
RaghavSood,
Ericson2314,
edolstra,
ajs124,
globin,
ulrikstrid,
maralorn,
FRidh,
alyssais,
Profpatsch,
sternenseemann,
dasJ,
Ma27,
fricklerhandwerk,
jonringer,
mkg20001,
aanderse,
piegamesde,
WilliButz,
infinisil,
cole-h,
figsoda,
IvarWithoutBones,
zowoq and
winterqt
|
It looks like you accidentally mass-pinged a bunch of people, which are now subscribed Please open a new pull request with your changes, link back to this one and ping the In order to avoid this in the future, there are instructions for how to properly |
Description of changes
Updated the default cudaPackages from 11.7 to 11.8 to address #222778. Pytorch doesn't support 12, yet, so this is an interim change. A nixpkgs-review with cudaSupport enabled is forthcoming.
Things done