-
-
Notifications
You must be signed in to change notification settings - Fork 17
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’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Is mkl-service still required? #69
Comments
@maresb removing dependencies is always a good thing. However, I have no better strategy to answer your question than recommending a series of tests:
The real place to improve the dependency management though is the Aesara feedstock: https://github.com/conda-forge/aesara-feedstock/ |
@michaelosthege thanks for the advice. I really have very little sense of what's going on internally. For example, the current release version is still 3.X, so I really don't understand what Aesara has to do with anything at the moment. Regarding the test suite, I just tried running
Do we need to add So... how shall we proceed? Shall we give up on 3.X and hope that 4.X comes out sometime soon, and then deal with dependencies there via Aesara? Or is there some subset of tests which I could run which are expected to pass on a vanilla install of PyMC3? |
The problems with installation of dependencies apply to both PyMC3 Ideally we should make both the Theano-PyMC & Aesara feedstocks more reliable. I guess one should start with the Aesara one, but the same fixes can probably just be copied over to the Theano-PyMC feedstock, getting us closer to a "LTS" situation.
AFAIK the max version requirement only applies to a subset of operations - one that's apparently covered by tests.
I'm seldom running the entire suite, but if you downgrade JAX, or pass |
Isn't JAX an optional dependency anyway? We also mark it as experimental so I don't see why we need to remove it after we advertised it, as it does work with the right JAX version. |
JAX is an implicit dependency via Conda doesn't really have the concept of optional dependency. |
Let's see how running the test suite in #70 goes. (At first without removing |
The failing test seems to me like a bug. Missing a version check perhaps? I'm not sure whether the bug is in PyMC3, Theano, or JAX... |
Any ideas about how to solve these issues? Note that I run purely Linux, so I am unable to reproduce anything locally.
|
I've never seen either of them before. Most likely an issue with the installed dependencies.. |
Update: After disabling the graphviz tests (all good, since graphviz is not a strict dependency, and the warning gives good installation instructions), the test suite now passes on Linux. ✔️ The Windows error which I mentioned above seems to be concurrent with the following error:
The previously reported error on OSX seems to have been transient. But now there are failures on all python versions with this message:
It seems to me likely that the warning text changed slightly, and now the test needs to be updated? I look forward to seeing what you come up with on Windows. |
Before running without mkl-service I want to get the test summary with the recommended install procedure. My steps:
I won't be at the computer again until late in the evening, but I'll post the full test summary as soon as I have one. |
@michaelosthege I'm sorry, this is embarrassing... After adding On my fork I just pulled master and merged it into my branch, so we'll see in a few hours if the tests complete in Windows. Sorry about that! |
@michaelosthege, even with m2w64-toolchain the result is the same. (See #70.) |
@maresb see above. Doesn't look like we can easily remove |
Thanks @michaelosthege for the test results!! Yes, I agree that it looks like for Windows we won't be able to remove Do you have any ideas about the failures in the tests which run on conda-forge CI? (#70) Namely this |
Adding We'll see in a few hours if the rest succeeds... |
Interesting... Windows with Python 3.8 is still running tests. Most of the tests have passed, but there are a few failures. On the other hand, Windows with Python 3.7 or Python 3.9 make it to the tests, but the tests fail rather catastrophically. https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=343481 Any ideas on why the behavior of Python 3.8 is so different from the other versions under Windows? |
@maresb which CI run are you referring to? |
I am and always have been referring to the conda-forge CI environment. In #70, I modified the meta.yaml to run the test suite. This should simulate what happens when someone with a fresh environment does For the results, please see: Thanks for the link to your issue, it seems like quite a rabbit hole. If I understand correctly, then the correct place to look for ideas on getting things to work is probably here on the If we get this feedstock sorted out, then the (non-developer) installation instructions for all platforms and all Python versions would simply be Is h5py somehow a base requirement, or is it optional? |
Yes.
When working on pymc-devs/pymc@576a733 I initially removed it, but eventually put it back in.
It would be even better to sort out the dependencies at the backend level (theano-pymc or aesara), because then the instructions could simplify to something like:
|
I ran
conda remove --force mkl mkl-service
and PyMC3 seemed to still work fine. Is it still a requirement?It's quite large, so it would be nice to get rid of.
The text was updated successfully, but these errors were encountered: