-
Notifications
You must be signed in to change notification settings - Fork 21.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.
Already on GitHub? Sign in to your account
Pillow v9.0.0 conda-forge package causes "OSError: [WinError 182] [...] error loading 鈥榗affe2_detectron_ops.dll'" #72293
Comments
Hey, sorry to hear that our packaging broke something. I added a bunch of bindings for 9.0.0 that pillow offers support for (like libimagequant), but that weren't previously built-in for the conda-forge packages. I'm surprised that these relatively minor changes seem to have broken pytorch. Do we know what |
This is caused by incompatible |
Thanks for the investigation! I'll remove libimagequant on windows for pillow 9.0.1 for the time being (before the larger issues around llvm-openmp vs. intel-openmp can be resolved) |
As a side note, don't forget to reinstall
|
should be fixed by conda-forge/pillow-feedstock#107 ? |
The issue with v9.0.0 seems to have been fixed in v9.0.1: pytorch/pytorch#72293 (comment)
Thanks for the reminder ping! I've tried pinning # Output from `conda list`
pillow 9.0.1 py38hd8e0db4_0 conda-forge # Identical test failures
C:\Miniconda\envs\ads_venv\lib\site-packages\ivadomed\inference.py:4: in <module>
import torch
C:\Miniconda\envs\ads_venv\lib\site-packages\torch\__init__.py:123: in <module>
raise err
E OSError: [WinError 182] The operating system cannot run %1. Error loading "C:\Miniconda\envs\ads_venv\lib\site-packages\torch\lib\caffe2_detectron_ops.dll" or one of its dependencies. CI run: https://github.com/neuropoly/axondeepseg/runs/5110622161?check_suite_focus=true EDIT: Just to be a little more thorough, here's the output of
Looks like |
@joshuacwnewton thanks I'm also getting the same issue still: dask/distributed#5775 (review) |
cc @h-vetinari |
Yeah, that was my bad, I changed the dependency specification only for the host and not runtime. Fix is here. Currently azure is unhappy for some reason in all of conda-forge, but once that is figured out, the new builds will be up in a matter of hours. |
I'll put an automerge label on it, so you'll even be able to retrigger it in my absence, by the posting magic invocation of |
Please note, that |
Small update: I tried re-running our CI, and managed to get a passing run (i.e. no OSError). The pillow version ( I presume the updated fix was quietly folded into the same patch version of pillow? (Please correct me if I'm wrong, as perhaps the fix came from somewhere else in the conda ecosystem.) Regardless, this issue seems to be resolved on our end. |
The issue with v9.0.0 seems to have been fixed in v9.0.1: pytorch/pytorch#72293 (comment)
When using Conda PyTorch version with CAFFE2 the libiomp5md.dll is not included in the torch\lib while its included in the pip package. So this issue is only happening when using Conda. To reproduce this:
|
That's because conda does not stuff non-python libraries into python-specific folders. That's a core part of the design, and intentionally different from pip (which has no concept of non-python dependencies).
This issue only happens when mixing pip with conda, or conda-forge with other channels (e.g. your |
@joshuacwnewton I suspect this is a conda issue since it's about a conflict between intel and llvm openmp. |
@mantaionut I'm not sure if I'm the best person to re-report? I'm just a downstream user, and admittedly I don't have a good understanding of the underlying problem here, since I moved on after the resolution from #72293 (comment). I also don't have any insight into e.g. why this was moved to "Blocked" on the PyTorch On Windows project board? So, it's hard for me to summarize what has happened here in the year since filing this issue. |
I've created an issue in conda ContinuumIO/anaconda-issues#13146. So I'll close this issue here. |
馃悰 Describe the bug
The conda-forge package for Pillow updated from 8.4.0 to 9.0.0 today (conda-forge/pillow-feedstock#104), and as a result we began receiving the following error when trying to
import torch
:(Sample CI run here.)
We can also confirm that pinning
Pillow<9.0.0
resolves the issue. (axondeepseg/axondeepseg@370c666 and corresponding CI run)Possibly related issues: #67145, #35803, discuss.pytorch.org#111995
Versions
Since the error happened in our CI, the quickest way for me to report versions is by copying and pasting the versions from that run. (Apologies that I can't easily run
collect_env.py
.)Output of conda info
We also diffed this output with the last passing run, and the majority of the packages versions are unchanged aside from Pillow.
Lastly, here is the Windows CI runner information:
cc @malfet @seemethere @peterjc123 @mszhanyi @skyline75489 @nbcsm
The text was updated successfully, but these errors were encountered: