Open jakirkham opened 2 months ago
An interesting wrinkle is dropping the /label/main
bit or use the -c
channel syntax with conda search
, it returns no packages found. For example
$ conda search --skip-flexible-search -c nvidia 'cuda[subdir=linux-aarch64]=12.5'
Loading channels: done
PackagesNotFoundError: The following packages are not available from current channels:
- cuda=12.5[subdir=linux-aarch64]
Current channels:
- https://conda.anaconda.org/nvidia/linux-aarch64
- https://conda.anaconda.org/nvidia/noarch
- https://repo.anaconda.com/pkgs/main/linux-aarch64
- https://repo.anaconda.com/pkgs/main/noarch
- https://repo.anaconda.com/pkgs/r/linux-aarch64
- https://repo.anaconda.com/pkgs/r/noarch
To search for alternate channels that may provide the conda package you're
looking for, navigate to
https://anaconda.org
and use the search bar at the top of the page.
Note: Flexible search is skipped otherwise we pick up cuda-python
, which is not what we are looking for here
cc @adibbley (for vis)
Maybe adding label/main bypasses the CDN, where it is missing?
That is what I'm thinking as well
For example, the same conda create
command above using nvidia/label/main
works
So sounds like this is just an issue mirroring to CDN
cc @chenghlee (for vis)
Huh. Is still an issue on your end? It seems to be working for me, so if you're still seeing the issue, very likely some weird CDN thing.
Yes am still seeing it unfortunately as is a colleague. Reproduced just before writing this
We tried a bunch of things:
Always the same issue with creation
Should add I wanted to confirm it wasn't something with corp networking. Have reproduced the same error on my personal laptop (which is x86_64
based)
Should add the same issue is not present with CUDA 12.4. So this works
conda create -c nvidia -n cuda12.4 cuda=12.4 --dry-run
Mentioning in case this provides any clues as to the differences between CUDA 12.4 & 12.5 packages
So I just tried downloading the repodata itself and am not seeing the packages there either
Commands:
curl -L https://conda.anaconda.org/nvidia/linux-aarch64/repodata.json > nvidia_linux-aarch64_repodata.json
cat nvidia_linux-aarch64_repodata.json | grep '12\.5\.0'
Output:
"cuda-python-12.5.0-py310hdae0e22_0.tar.bz2": {
"version": "12.5.0"
"cuda-python-12.5.0-py311hdae0e22_0.tar.bz2": {
"version": "12.5.0"
"cuda-python-12.5.0-py312hdae0e22_0.tar.bz2": {
"version": "12.5.0"
"cuda-python-12.5.0-py39hdae0e22_0.tar.bz2": {
"version": "12.5.0"
For some reason only cuda-python
shows up, which was uploaded before any of the CTK packages were uploaded
Can you try an anaconda login
(if you have an anaconda.org) and then attempt the conda create
command again?
This has been escalated to Anaconda's infrastructure team. We have identified two possible causes and are actively investigating them.
Thanks Cheng! 🙏
Hope everyone had a good weekend!
Am curious how things are going here
It looks like the channel cloning CDN is having problems dealing with .conda format files on the nvidia
channel. We're going to deploy an attempted fix in the morning EU time.
Ah ok. That makes sense. This is our first release using .conda
Thanks Cheng! 🙏
Please let us know when we should recheck
Did this fix get deployed?
We tried it and it didn't work. 😞 We have an alternative approach, but that will take longer to roll out. (Still working on getting an estimate.)
In the meantime, you might want to consider falling back to the older .tar.bz2 format, at least for this release until we get the CDN sorted out.
Gotcha, thanks Cheng! 🙏
If we uploaded .tar.bz2
and kept .conda
packages, would that work or would that cause issues? IOW should we delete the .conda
packages and then upload .tar.bz2
packages?
You can upload both formats (i.e., keep the .conda
ones already there). The .org backend definitely supports that, and I don't think having both formats would cause any problems since bug in the CDN cloning process is [essentially] just ignoring the repodata "packages.conda"
entries supplied by the backend.
BTW, the reason -c nvidia/label/main
made the packages appear is that adding labels to the channel name bypasses the CDN clones and connects directly to the .org backend.
We reuploaded the CUDA packages in .tar.bz2 format and things seem to be working as expected. Thanks!
Thanks Cheng and Alex! 🙏
Have rewritten the OP and title to point out the remaining issue of mirroring nvidia
channel .conda
packages to CDN
Am curious what the current status is on this issue. It would be nice to have .conda
packages for future CUDA releases
Expect users would appreciate the smaller packages for downloads
Not to mention having access to package metadata without needing to do a full decompression is more noticeable with these larger packages
This happened again with CUDA 12.5.1. We used .tar.bz2
Recently we uploaded
.conda
packages for CUDA 12.5 to thenvidia
channel. However we discovered they were not mirroring to CDN.We noticed this by trying to install CUDA 12.5 from the
nvidia
channel, but got the following error:However I can see the package is available via
conda search
Also via Anaconda.org
Since these were uploaded, we were recommended to upload
.tar.bz2
packages to thenvidia
channel, which the CDN appears to handle betterStill it would be good to support
.conda
packages on thenvidia
channel CDN too. This issue remains open to address that