pytorch-lightning
pytorch-lightning copied to clipboard
Lightning-AI
Conda installation defaults to using a very old version (0.8.5) instead of latest version (1.7)
🐛 Bug
On a new environment, doing the conda installation results in installing "conda-forge/noarch::pytorch-lightning-0.8.5-py_0".
To Reproduce
conda create -n test_env python=3.9
conda install ipython jupyter
conda install pytorch torchvision torchaudio cudatoolkit=11.6 -c pytorch -c conda-forge
conda install pytorch-lightning -c conda-forge
Expected behavior
It should install latest version when possible.
Environment
- PyTorch Version (e.g., 1.10): 1.12
- Python version (e.g., 3.9): 3.9
- OS (e.g., Linux): Linux
- CUDA/cuDNN version: 11.6
- GPU models and configuration: 3090
- How you installed PyTorch (
conda,pip, source): conda
I was able to reproduce it. It is enough to do:
conda create -n test_env python=3.9
conda install pytorch-lightning -c conda-forge
@Borda do you know what could be the reason for this? The latest version is correctly hosted here.
do you know what could be the reason for this? The latest version is correctly hosted here.
not sure, but it sounds strange that would take a version so old... :( also I think that back then when 0.8.5 was released py 3.9 was not out yet :rabbit:
Hi @Borda , do you have any updates on this issue? I have the exact same issue, and trying in a fresh conda environment - like this link suggests - does not resolve it.
Hi everyone, what conda version are you using? I just downloaded conda 4.14.0 and cannot reproduce on Linux
I am running in circles while trying to reproduce this and was just not able to. Only thing that is remotely notable about pytorch-lightning 0.8.5 is the fact that it is the last version allowing tensorboard<2.
Could please all of you who are able to reproduce this behavior (@awaelchli, @maxencewynen, @TheShadow29) post here:
- Series of commands leading to creating such environment, ideally as minimal as possible.
-
conda listof the resulting environment. -
conda conda list -n base - your OS and ideally also method of obtaining conda as it seems different distributions differ in some default options.
Thank you!
There's one more notable difference between pytorch-lightning=0.8.5 and any newer version - pytorch-lightning=0.8.5 has a lower bound version pin on future>=0.17.1, whereas newer versions require future without any pin all the way to 1.7.2 and from 1.7.3 we don't require future at all. Could it be it? At the time of reporting this issue, 1.7.2 was the latest version published on conda-forge and 1.7.3 wasn't released out yet.
Series of commands leading to creating such environment, ideally as minimal as possible.
I initially reproduced using https://github.com/Lightning-AI/lightning/issues/14374#issuecomment-1225849005 But now, spookingly, I CANNOT anymore. How is it possible. Still same conda version 4.12 (miniconda). And I didn't change anything since last time.
I can confirm as well. It now uses 1.7.6. I am not sure what changed though. My conda env is same as it was last time. My guess is some dependency somewhere was removed but I am not sure.
It might have been future, as I said, this dependency was dropped between 1.7.2 and 1.7.3 (at least on conda-forge side). Both of you reported at the time when latest version out was 1.7.2. It could be the fact that it had hard time finding some.
Let's wait on @maxencewynen and if he reports no issue anymore, I will close this. If he's able to still reproduce, I will investigate further (although I don't know where to anymore :joy:)
It is possible something was broken and got fixed by this PR https://github.com/conda-forge/pytorch-lightning-feedstock/pull/97
@maxencewynen Has this been resolved for you as well? Would you mind checking that it downloads the latest version now? Thanks!
Faced this today with latest conda (22.9.0). Resolved only with specifying pytorch-lightning==1.7.7. For some reasons when installing pl 1.7.7 conda also downgraded openssl from 3.0.5 all the way down to 1.1.1, could it be connected? Minimal steps to reproduce (or not?): conda create -n test_pl python=3.10 conda activate test_pl conda install pytorch-lightning
(probably worth mentioning that I have conda-forge added to default channels in .condarc)
Also observed this happening (conda install pytorch-lightning -c conda-forge installed pytorch-lightning 0.8.5 and openssl 1.1.1).
Edit: conda update pytorch-lightning -c conda-forge produced a very large number of packages to update, including updating pytorch-lightning to 1.7.7.
