mlx: [BUG] bus error

Thanks @awni, then unless there’s some unlikely difference between M3 Pro and Max I guess this is a local issue. I’ll continue to explore what’s the cause and post any solution in case someone else faces the same problem at some point.

There is a simple test case in this thread. Maybe if you have a second you can play around with it and see if it also gives you a bus error. That would strongly suggest something is borked in your environment.

That actually works fine for me.

For reproducibility: I tried compiling via

/usr/bin/clang++ -shared -stdlib=libc++ -std=c++17 -undefined dynamic_lookup $(python3 -m pybind11 --includes) test.cpp -o test$(python3-config --extension-suffix)

and

c++ -shared -std=c++17 -undefined dynamic_lookup $(python3 -m pybind11 --includes) test.cpp -o test$(python3-config --extension-suffix)

And I always get the correct output when running the python test.

Note: with mlx I get a bus error not a segmentation fault.

There is a simple test case in this thread. Maybe if you have a second you can play around with it and see if it also gives you a bus error. That would strongly suggest something is borked in your environment.

Thanks @gboduljak. I’m admittedly a little stumped by this one. Would be good to check on an M3 Pro to verify that has nothing to do with it… (@jagrit06 might be able to help there).

I get this: ValueError: [tril] array must be at least 2-D

I am also getting this error. I am working on M2.

FWIW this is what I get when running make test

Test project <path_to_mlx>/build
    Start 1: tests
1/1 Test #1: tests ............................Bus error***Exception:   0.04 sec

0% tests passed, 1 tests failed out of 1

Total Test time (real) =   0.05 sec

The following tests FAILED:
	  1 - tests (Bus error)
Errors while running CTest
Output from these tests are in: <path_to_mlx>/build/Testing/Temporary/LastTest.log
Use "--rerun-failed --output-on-failure" to re-run the failed cases verbosely.
make: *** [test] Error 8

But you see the same problem from C++ only right? So it seems unlikely to be a binding issue 🤔

Yes that’s the weird part.

Can you also see the output of:

• uname -m

arm64

• How did you install cmake? Via brew which cmake?

via brew: /opt/homebrew/bin/cmake

Is it just that case? Can you do mx.tril(mx.ones((10, 10)))?

That works actually! Looks like the problem may be with raising exceptions/throwing errors - the test I pointed out above is the first one to expect an exception.

So the tril op seems to consistently cause a bus error. I tested it manually and also all the tests up to here work https://github.com/ml-explore/mlx/blob/449b43762e3f970576f054e54066123c0f37246e/python/tests/test_ops.py#L331 this one is the first to cause a bus error (in test_ops.py).

Simply running

import mlx.core as mx

mx.tril(mx.zeros([1]))

results in a bus error.