Adding Jax tests for the M1 mac. #508

alxmrs · 2024-07-20T20:50:04Z

I'm beginning to explore #304 in greater depth. Since the only local GPU I have access to is an M1 chip (I have an M1 Macbook Air), I thought I would replicate this environment in CI.

alxmrs · 2024-07-20T20:57:49Z

Locally, I'm hitting a large number of errors. It looks like jax-metal is still highly experimental.

104 failed tests!

FAILED cubed/tests/test_array_api.py::test_ones[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_ones_like[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_add_top_level_namespace[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_add_different_chunks[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_add_different_chunks_fail[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_matmul[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: /Users/alxrsngrtn/git/cubed/cubed/array_api/linear_algebra_functions.py:63:12: error: 'mps.matmul' op operand...
FAILED cubed/tests/test_array_api.py::test_broadcast_arrays[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_all_zero_dimension[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_ones[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_ones_like[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_add_top_level_namespace[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_add_different_chunks[processes] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-6
FAILED cubed/tests/test_array_api.py::test_add_different_chunks_fail[processes] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-7
FAILED cubed/tests/test_array_api.py::test_matmul[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: /Users/alxrsngrtn/git/cubed/cubed/array_api/linear_algebra_functions.py:63:12: error: 'mps.matmul' op operand...
FAILED cubed/tests/test_array_api.py::test_outer[processes] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-10
FAILED cubed/tests/test_array_api.py::test_broadcast_arrays[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_all_zero_dimension[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_eye[-1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_eye[0] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_eye[1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_linspace[True] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_linspace[False] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_tril_triu[-1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_tril_triu[0] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_tril_triu[1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_index_2d_step[shape0-chunks0-ind0-new_chunks_expected0] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_array_api.py::test_tensordot[1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: /Users/alxrsngrtn/git/cubed/cubed/array_api/linear_algebra_functions.py:156:8: error: 'mps.matmul' op operand...
FAILED cubed/tests/test_array_api.py::test_tensordot[axes1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: /Users/alxrsngrtn/git/cubed/cubed/array_api/linear_algebra_functions.py:156:8: error: 'mps.matmul' op operand...
FAILED cubed/tests/test_core.py::test_compute_is_idempotent[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_default_spec[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_array_pickle[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: /Users/alxrsngrtn/git/cubed/cubed/array_api/linear_algebra_functions.py:63:12: error: 'mps.matmul' op operand...
FAILED cubed/tests/test_core.py::test_compute_is_idempotent[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_default_spec[processes] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-42
FAILED cubed/tests/test_core.py::test_array_pickle[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: /Users/alxrsngrtn/git/cubed/cubed/array_api/linear_algebra_functions.py:63:12: error: 'mps.matmul' op operand...
FAILED cubed/tests/test_core.py::test_map_blocks_with_different_block_shapes - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-45
FAILED cubed/tests/test_core.py::test_rechunk_intermediate - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_default_spec_config_override - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_partial_reduce - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_visualize - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_core.py::test_plan_scaling[10] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[20] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[50] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[100] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[200] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[500] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[1000] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[2000] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_core.py::test_plan_scaling[5000] - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_executor_features.py::test_compute_arrays_in_parallel[single-threaded-True] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_executor_features.py::test_compute_arrays_in_parallel[single-threaded-False] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_executor_features.py::test_compute_arrays_in_parallel[threads-True] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_executor_features.py::test_compute_arrays_in_parallel[threads-False] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_executor_features.py::test_compute_arrays_in_parallel[processes-True] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_executor_features.py::test_compute_arrays_in_parallel[processes-False] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-53
FAILED cubed/tests/test_gufunc.py::test_apply_gufunc_axes_two_kept_coredims - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_nan_functions.py::test_nansum - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_nan_functions.py::test_nansum_allnan - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a0-chunks_b0-expected_chunksize0] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a1-chunks_b1-expected_chunksize1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a2-chunks_b2-expected_chunksize2] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a3-chunks_b3-expected_chunksize3] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a4-chunks_b4-expected_chunksize4] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a5-chunks_b5-expected_chunksize5] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise[chunks_a6-chunks_b6-expected_chunksize6] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise_2d[chunks_a0-chunks_b0-expected_chunksize0] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise_2d[chunks_a1-chunks_b1-expected_chunksize1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise_2d[chunks_a2-chunks_b2-expected_chunksize2] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise_2d[chunks_a3-chunks_b3-expected_chunksize3] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_elemwise_2d[chunks_a4-chunks_b4-expected_chunksize4] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_blockwise_2d[chunks_a0-chunks_b0-expected_chunksize0] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_blockwise_2d[chunks_a1-chunks_b1-expected_chunksize1] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_blockwise_2d[chunks_a2-chunks_b2-expected_chunksize2] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_blockwise_2d[chunks_a3-chunks_b3-expected_chunksize3] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_blockwise_2d[chunks_a4-chunks_b4-expected_chunksize4] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_broadcast_scalar - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_ops.py::test_unify_chunks_broadcast_2d - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_optimization.py::test_no_fusion - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_optimization.py::test_no_fusion_multiple_edges - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_optimization.py::test_fuse_unary_op - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_binary_op - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_unary_and_binary_op - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_mixed_levels - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_diamond - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_mixed_levels_and_diamond - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_mixed_levels_and_diamond_complex - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_repeated_argument - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_other_dependents - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_unary_large_fan_in - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_large_fan_in_default - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_large_fan_in_override - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_with_merge_chunks_unary - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_with_merge_chunks_binary - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_merge_chunks_unary - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_merge_chunks_binary - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_partial_reduce_unary - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_partial_reduce_binary - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_fuse_only_optimize_dag - NotADirectoryError: [Errno 20] Not a directory: 'dot'
FAILED cubed/tests/test_optimization.py::test_optimize_stack - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_random.py::test_random[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_random.py::test_random_add[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_random.py::test_random_seed[single-threaded] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_random.py::test_random[processes] - jaxlib.xla_extension.XlaRuntimeError: UNKNOWN: <unknown>:0: error: 'func.func' op One or more function input/output data types are not supported.
FAILED cubed/tests/test_random.py::test_random_add[processes] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-55
FAILED cubed/tests/test_random.py::test_random_seed[processes] - pytest.PytestUnhandledThreadExceptionWarning: Exception in thread Thread-56
======================================= 104 failed, 258 passed, 18 skipped, 32 deselected, 1 warning in 208.40s (0:03:28)

alxmrs · 2024-07-21T16:11:29Z

I've pushed some changes to cut the failed tests down in half locally. I'll definitely need design opinions on my review. The next thing I plan to tackle is randomness, which is a special case for Jax and all GPU acceleration.

tomwhite

Thanks for working on this @alxmrs. I'm honoured that you are working on Cubed during your round the world trip!

I have a Mac mini M1 so I should be able to try it when you've got it working. Can you post your environment versions and any pointers to getting it set up please?

.github/workflows/jax-tests.yml

cubed/array_api/creation_functions.py

alxmrs · 2024-07-22T14:25:57Z

Can you post your environment versions and any pointers to getting it set up please?

I’m on mobile at the moment, but in the meantime: I’m using Python 3.11 (compiled for ARM). I’ve followed these instructions to set up jax for the M1 (which specifically means installing jax-metal):

https://developer.apple.com/metal/jax/

Right now, I’m trying out rye instead of using the usual conda. It’s going good so far (it takes getting used to their opinions). It fits well with Apple’s instructions, which say to create a standard python virtualenv.

alxmrs · 2024-07-23T13:07:56Z

I'm hitting the same errors faced in #494, namely that "dot" is not found during visualize.

NotADirectoryError: [Errno 20] Not a directory: 'dot'

Created an affordance to apply a jit to blockwise functions after operator fusion. This will let the user better use accelerators. cubed-dev#508 needs to be merged first.

tomwhite

This is looking good. I added a few comments about dtype handling.

The CI workflow is crashing with a seg fault - do you see the same on your machine?

tomwhite · 2024-07-24T08:14:03Z

cubed/array_api/creation_functions.py

        dtype = nxp.arange(start, stop, step * num if num else step).dtype
+        for k, dtype_ in default_dtypes(device=device).items():
+            if nxp.isdtype(dtype, k):
+                dtype = dtype_


It's not clear to me what this is doing or why it is needed. If nxp is the jax namespace doesn't the call to arange already return the correct dtype (int32) - or does jax metal just return int64 or fail?

It might help to factor out a function to do this (given it is duplicated below too) with a name describing what it does, and perhaps a comment too.

or does jax metal just return int64 or fail?

Yes, it looks like jax metal returns 64 bit precisions and fails; this ensures the correct precision.

I've factored this out to a function, good shout.

cubed/array_api/creation_functions.py

cubed/tests/test_array_api.py

tomwhite · 2024-07-24T08:23:51Z

cubed/tests/test_array_api.py

+    x = np.arange(400, dtype=np.float32).reshape((20, 20))
+    a = xp.asarray(x, chunks=(5, 4), dtype=xp.float32)
+    y = np.arange(200, dtype=np.float32).reshape((20, 10))
+    b = xp.asarray(y, chunks=(4, 5), dtype=xp.float32)


Ditto - we need to test that leaving out dtype works OK in these cases.

alxmrs · 2024-07-24T10:42:04Z

Thanks for the review! I’ll address these when I can.

The CI workflow is crashing with a seg fault - do you see the same on your machine?

No! Tests are running without any faults locally for me. I’ll do some digging to better understand the CI environment.

alxmrs · 2024-07-27T21:38:32Z

@tomwhite It looks like to use the GPU on the M1 actions, we need to enable a premium action runner: https://github.blog/news-insights/product-news/introducing-the-new-apple-silicon-powered-m1-macos-larger-runner-for-github-actions/

I think this is the cause of the segfault (from what I can tell from related discussions): pytorch/pytorch#111449 (comment).

How do you think we should proceed here? Should we attempt to have CI target GPUs, or should I configure Jax to run on the CPU?

tomwhite · 2024-07-29T08:26:23Z

How do you think we should proceed here? Should we attempt to have CI target GPUs, or should I configure Jax to run on the CPU?

We should certainly have JAX running against the CPU in CI as it tests that Cubed works with the JAX array API. For testing on Mac M1 GPUs, I think we can add that later - particularly since it's a paid for option? It would be good to get the work that you have done here merged, so if you change the CI back then I'm happy to merge it.

BTW I tried installing JAX metal on my Mac Mini M1 to run the tests, but I got an error (SystemError: nanobind::detail::nb_func_error_except(): exception could not be translated!). It looks like I need to update the OS to Sonoma (I'm on Ventura), so I'll try to do that at some point - but this shouldn't block merging since you've got it working.

TomNicholas · 2024-07-29T19:13:59Z

particularly since it's a paid for option?

There was some stuff at SciPy about quansight being able to give out free access to NVIDIA GPUs for scientific python projects to use in CI. But this seems like a much later concern only for once it works on CPU and seems useful.

tomwhite · 2024-08-02T09:10:42Z

Looks like this is getting close @alxmrs. There are still a few places where you've changed the tests to have a lower precision dtype, where we should also test that it works when it's left at the default. Can we merge once they are resolved?

alxmrs · 2024-08-03T16:22:18Z

That sounds good to me Tom. Since my development time is sporadic and limited, I'll try to make the Jax features I work on independent from each other from here on out. Today, I extracted #536; since this is a bigger / flakier PR, I'll probably need a few more sessions to get it to fully land.

I'm extracting cubed-dev#508 into smaller bites.

I'm extracting #508 into smaller bites.

alxmrs · 2024-08-12T10:45:38Z

Hey Tom! I should have mentioned this earlier. Can you run the workflows again? I think this PR is ready.

tomwhite · 2024-08-12T11:14:43Z

The Dask test failure is a flaky test (#549), but the Array API test failures look like they are real.

Inlined in order to not violate API boundaries. Trying to put this in a good place ends up leading to a circular import issue.

tomwhite · 2024-09-23T17:03:50Z

@alxmrs thanks for pushing this forward! Do you think it might make things easier to reduce the scope by e.g. targeting JAX on CPU and focusing on the device/dtype inspection stuff.

Or maybe there's another way of splitting things up? I'd be happy to merge smaller PRs!

alxmrs · 2024-09-24T01:07:34Z

I totally agree — at this point, it’s best for me to split up this large CL. Yeah, targeting JAX on CPU while respecting default dtypes is a good idea for a first cut. ETA: I’ll start on this after focusing on landing std/var.

…

On Tue, Sep 24, 2024 at 2:04 AM Tom White ***@***.***> wrote: @alxmrs <https://github.com/alxmrs> thanks for pushing this forward! Do you think it might make things easier to reduce the scope by e.g. targeting JAX on CPU and focusing on the device/dtype inspection stuff <https://data-apis.org/array-api/latest/API_specification/inspection.html> . Or maybe there's another way of splitting things up? I'd be happy to merge smaller PRs! — Reply to this email directly, view it on GitHub <#508 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AARXAB5K6PAZKXRODOMPXQ3ZYBCY3AVCNFSM6AAAAABLGHA23GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNRYHA3DMNJUGI> . You are receiving this because you were mentioned.Message ID: ***@***.***>

tomwhite · 2024-09-24T08:45:58Z

Great - thanks @alxmrs!

tomwhite reviewed Jul 22, 2024

View reviewed changes

.github/workflows/jax-tests.yml Show resolved Hide resolved

cubed/array_api/creation_functions.py Outdated Show resolved Hide resolved

cubed/array_api/creation_functions.py Outdated Show resolved Hide resolved

tomwhite mentioned this pull request Jul 22, 2024

Single machine tracking issue #514

Open

9 tasks

alxmrs marked this pull request as ready for review July 23, 2024 13:39

alxmrs mentioned this pull request Jul 23, 2024

Adding Jit functionality to Plan finialization. alxmrs/cubed#1

Open

tomwhite reviewed Jul 24, 2024

View reviewed changes

alxmrs force-pushed the m1-jax branch from fcfc56d to 2ff5158 Compare July 25, 2024 15:24

alxmrs force-pushed the m1-jax branch 2 times, most recently from e78cd21 to 2f7c324 Compare July 31, 2024 19:52

alxmrs mentioned this pull request Aug 3, 2024

Adding compile_function as execute option. #536

Merged

alxmrs added a commit to alxmrs/cubed that referenced this pull request Aug 3, 2024

Turning on JAX tests for all PRs.

62b135b

I'm extracting cubed-dev#508 into smaller bites.

This was referenced Aug 3, 2024

Turning on JAX tests for all PRs. #537

Merged

Turning on JAX tests for all PRs. alxmrs/cubed#3

Closed

alxmrs force-pushed the m1-jax branch from 9a7a419 to 3637e23 Compare August 3, 2024 16:51

tomwhite pushed a commit that referenced this pull request Aug 4, 2024

Turning on JAX tests for all PRs. (#537)

819f6c8

I'm extracting #508 into smaller bites.

alxmrs added 25 commits September 21, 2024 18:19

Cleanup for PR.

896a738

Added developer instructions for getting set up on the M1+ processor.

a0080a9

Fix env variable expressions.

5b00d82

Looks like 14 specifically supports python 3.11 and up.

e9b53e3

Splitting install steps for jax for mac and non-mac.

1ed8ac2

Mac install now enables jaxlib compatibility.

72bd026

Put the mac-specific env setting in the right place.

8b6b1f4

Test the xlarge runner.

c4b51ad

Extracted to method _to_default_precision.

7566849

Added helpful note.

2f470d5

No need to cast in the linspace test.

06e030d

Remove macos-specific tests.

e642f8d

Back tracking to previous macos build

2afba45

Mac should use CPU Jax

38831b6

Remove jax mac extra install.

9bd0210

Tensordot testing with and without hardcoded types.

18e0aa0

Defaulting to 64 bit precision.

4d132ff

rm unneeded test arg.

9f6017a

Revert tensordot change.

1449008

Revert hardcoding precision as much as possible.

38dfc32

Make precision a focus point in test matrix.

aebb2ab

Fix underlying bug in asarray.

4687d33

Exposing to_default_precision as a backend_array_api.

92a31cc

All cubed tests pass!

dbefff4

Inlined extracted function.

b8ce6ec

Inlined in order to not violate API boundaries. Trying to put this in a good place ends up leading to a circular import issue.

alxmrs force-pushed the m1-jax branch from e41e6fa to b8ce6ec Compare September 21, 2024 09:20

tomwhite mentioned this pull request Oct 5, 2024

Add array API inspection utilities #592

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adding Jax tests for the M1 mac. #508

Adding Jax tests for the M1 mac. #508

alxmrs commented Jul 20, 2024

alxmrs commented Jul 20, 2024 •

edited

Loading

alxmrs commented Jul 21, 2024

tomwhite left a comment

alxmrs commented Jul 22, 2024

alxmrs commented Jul 23, 2024

tomwhite left a comment

tomwhite Jul 24, 2024

alxmrs Jul 30, 2024 •

edited

Loading

tomwhite Jul 24, 2024

alxmrs commented Jul 24, 2024

alxmrs commented Jul 27, 2024

tomwhite commented Jul 29, 2024

TomNicholas commented Jul 29, 2024

tomwhite commented Aug 2, 2024

alxmrs commented Aug 3, 2024

alxmrs commented Aug 12, 2024

tomwhite commented Aug 12, 2024

tomwhite commented Sep 23, 2024

alxmrs commented Sep 24, 2024 via email

tomwhite commented Sep 24, 2024

Adding Jax tests for the M1 mac. #508

Are you sure you want to change the base?

Adding Jax tests for the M1 mac. #508

Conversation

alxmrs commented Jul 20, 2024

alxmrs commented Jul 20, 2024 • edited Loading

alxmrs commented Jul 21, 2024

tomwhite left a comment

Choose a reason for hiding this comment

alxmrs commented Jul 22, 2024

alxmrs commented Jul 23, 2024

tomwhite left a comment

Choose a reason for hiding this comment

tomwhite Jul 24, 2024

Choose a reason for hiding this comment

alxmrs Jul 30, 2024 • edited Loading

Choose a reason for hiding this comment

tomwhite Jul 24, 2024

Choose a reason for hiding this comment

alxmrs commented Jul 24, 2024

alxmrs commented Jul 27, 2024

tomwhite commented Jul 29, 2024

TomNicholas commented Jul 29, 2024

tomwhite commented Aug 2, 2024

alxmrs commented Aug 3, 2024

alxmrs commented Aug 12, 2024

tomwhite commented Aug 12, 2024

tomwhite commented Sep 23, 2024

alxmrs commented Sep 24, 2024 via email

tomwhite commented Sep 24, 2024

alxmrs commented Jul 20, 2024 •

edited

Loading

alxmrs Jul 30, 2024 •

edited

Loading