Testing of conda-forge package for ambertools24.8

[dacase]

Not sure if a new issue is appropriate here, but the corresponding PR's have been closed.

I installed linux-64/ambertools-24.8-cuda_None_nompi_py310h1ac786c_100.conda, which was the first one for PR #162 to get posted. (I see that there are now a lot more.)

Here are the results of the serial test suite:

1774 file comparisons passed
2 file comparisons failed (2 of which can be ignored)
8 tests experienced errors

Overall, this is pretty promising.

Errors: the sander API for C,C++ and Fortran fail, presumably because the conda compilers used for building the sander library are not included in the "run" section of meta.yaml. My opinion is that the sander API should not be supported in a conda-forge package. Users that want to link libsander to their own driver programs presumably should be building from source -- there is no simple other way (that I can see) to ensure that the same compilers are used to build the library as to build the driver code. This is particularly a problem for the Fortran API, since different versions of gfortran produce module files that are not compatible with each other.

The only other error is in the pytraj test...sigh. It's late tonight and I don't have time to test this further; but it's possible that pytraj actually works, and it's the test case that is failing. (Something like this has happened before.)

Thanks to @mikemhenry @mattwthompson and all the others who have worked on this! If
there is interest, I can post the exact steps to run the "official" AmberTools test suite.

I'll try out some parallel tests next, but I'm not sure how soon I can get to it. If you have requests for certain versions to test, let me know. I'm pretty clueless about why there are so many versions of linux, no cuda, nompi, and how "ordinary" users know which one to install.

[xiki-tempula]

I noted that it is pulling the source code of AmberTools24_rc5.tar.bz2. Is this correct? Should you be pulling down an RC version?
https://github.com/conda-forge/ambertools-feedstock/blob/main/recipe/meta.yaml#L21

[dacase]

Yes, the above is the correct source code to pull down. RC5 is the final released version. We use a slightly obfucated file name to encourage most users to download via the link on the Amber web page, so that we can maintain a (somewhat) useful idea of the number of downloads.

[mikemhenry]

If there is interest, I can post the exact steps to run the "official" AmberTools test suite.

I would be! We can run the tests after we do the build to automate them.

Also now that the builds have been uploaded onto conda-forge proper, you should be able to install the package like micromamba install -c conda-forge -n ambertools24-test ambertools=24

It may help with getting the tests to pass since it will ensure that it picks out the correct libs the package was compiled against (sometimes when installing the .conda manually it can have issues resolving things correctly).

[dacase]

Test procedure should look something like this, starting in some pretty arbitrary folder:

wget https://ambermd.org/downloads/AmberTools24_rc5.tar.bz2
tar xvf AmberTools24_rc5.tar.bz2
cd amber24_src
rsync -av test $CONDA_PREFIX # assumes $CONDA_PREFIX points to where ambertools is installed
cp cmake-packaging/top-level-test-makefile $CONDA_PREFIX/Makefile
cd AmberTools
rsync -av test $CONDA_PREFIX/AmberTools
cd $CONDA_PREFIX

source amber.sh
make test.serial
export DO_PARALLEL='mpirun -np 4' # if MPI is enabled
make test.parallel
unset DO_PARALLEL
make test.cuda.serial # if CUDA is enabled

The test targets will point to where the summaries and log files are located.

[In principle, there is also a test.cuda.parallel target, but AmberTools has essentially no parallel-GPU components.]

[dacase]

Automating this test suite is more complex. Basically, our CI process requires that all tests pass on one particular Linux environment, which is not at the cutting edge: I need to check, but I think it uses gcc11, nvcc 11.8; not sure right now which MPI stack is used.

Even going to a different Linux environment can occasionally trigger errors, and OSX (either osx-64 or osx-arm64) can also show failures. We manually test release candidates in about a dozen environments, and make decisions on whether test case failures really represent problems for users. [FWIW, AmberTools24 was not tested very much on osx-arm64, as you probably know.]

The conda-forge Linux package I tested (above) is really quite good: the API failures can either be addressed by adding compilers to the run: section of meta.yaml, or by deciding that a pre-built package should not be expected to support that facility (which I think is rarely used.)

The pytraj failure is being quite actively investigated by its author, Hai Nguyen. It is a very complex python project, compiling python extensions from C++ code, and using cython to boot. I hope we can get things fixed by AmberTools25. I'm hoping that openFF and other potential users can live with a disabled pytraj. (I never use that myself.)

[mikemhenry]

Awesome, thank you for that!

Good to know the errors are not mission-critical.

"We" (the openfe people) mostly use antechaber and sqm since a lot of our partners like to use GAFF, which is why I didn't encounter some of the pytraj issues.

pytraj is pretty complex, if Hai Nguyen ever wants to just test to see if a new version/bug fix will work on conda-forge, I can help with getting a "test" build that people can install using a tag, so it won't get uploaded to the main channel but an rc or beta channel so users will have to opt in to test it.

[hainm]

The only other error is in the pytraj test...sigh. It's late tonight and I don't have time to test this further; but it's possible that pytraj actually works, and it's the test case that is failing. (Something like this has happened before.)

@dacase I've done mamba install -c conda-forge ambertools=24 with python 3.10 on my macbook + linux desktop. I ran some pytraj's functional tests and it seems working fine. So it's working. Cheers.

Thanks everyone for taking care of this.

[dacase]

I ran some pytraj's functional tests and it seems working fine. So it's working. Cheers.

Thanks. Can you give details of one or more tests that you ran, so that I can try this myself?
Among other things, I'd like to fix the pytraj test case (which is segfaulting for me) to more correctly represent the state of the pytraj installation.

[hainm]

I ran some pytraj's functional tests and it seems working fine. So it's working. Cheers.

Thanks. Can you give details of one or more tests that you ran, so that I can try this myself? Among other things, I'd like to fix the pytraj test case (which is segfaulting for me) to more correctly represent the state of the pytraj installation.

@dacase I would do

pip3 install mock # if not yet
pip3 install pytest # if not yet
git clone https://github.com/Amber-MD/pytraj
cd pytraj
git clone https://github.com/Amber-MD/cpptraj
cd cpptraj
export CPPTRAJHOME=`pwd`
cd ../tests
python3 -m pytest test_analysis/ -vs

This covers most of the public functionalities of pytraj and most of the tests should pass.
Cheers

[dacase]

Does anyone know the syntax for 'conda install' (or 'micromamba install') to get ambertools 24 installed with MPI turned on? Just asking for ambertools=24 gets a nompi version. I've more or less blindly tried various methods to select openmpi (say), but they all report "project not found".

"conda search" yields the names of all the projects, but there is some syntax I'm missing about how to actually install a particular one.

[xiki-tempula]

conda install "ambertools=24.8=*openmpi*"

[dacase]

Thanks! It was the second "=" sign that I was missing.