Lisandro Dalcin

Results 291 comments of Lisandro Dalcin
trafficstars

Yes, but it could be that previous tests have a side effect that end up breaking the dup+free. Remember that mpi4py tests run within a single invocation of mpiexec. And...

Oh, hold on. My test uses the world comm.

> I modified the test to use MPI_COMM_WORLD and still cannot reproduce the problem that you are hitting. I'll have to bisect all the changes by hand using GitHub Actions....

@hppritcha I believe I got confused. The issue actually happens with `MPI_Comm_create_from_group` using the main branch (see the [[logs]](https://github.com/mpi4py/mpi4py-testing/actions/runs/3158087627/jobs/5139722238#step:21:2805)), I paste the output below: ``` ====================================================================== ERROR: testCreateFromGroup (test_comm.TestCommSelf) ----------------------------------------------------------------------...

@hppritcha After disabling tests for `MPI_Comm_create_from_group`, all tests are [green](https://github.com/mpi4py/mpi4py-testing/actions/runs/3173401013/jobs/5168974421) with this PR. This is good to merge from my side. However, we now have this new issue of `MPI_Comm_create_from_group`...

> looks like the python test is tripping up on the comm compare Good catch! This mean that the way I wrote the test is not proper. I fixed it....

>> When running in singleton init mode, MPI_Group_from_session_pset now fails with error class MPI_ERR_OTHER. Perhaps MPI_ERR_UNSUPPORTED_OPERATION would be more appropriate and in line with other recent changes from you? >...

@rgommers What about supporting the thing if some environment variable is set? [Previous practice](https://numpy.org/doc/stable/reference/global_state.html#interoperability-related-options) would support my proposal. Additional mechanisms could be added, but an environment variable is a good...

@seberg I'm 100% on your side regarding the need to find a proper solution. My environ variable proposal is just a way to strength NumPy's statement of need and provide...

> I think if we add one now, we can only remove it once we either have DLPack support for a readonly flag, or we have moved to unconditionally exporting....