TST: run tests on CPU+GPU #221
Conversation
| @@ -298,7 +300,7 @@ def _op( | |||
| and idx.dtype == xp.bool | |||
| and idx.shape == x.shape | |||
| ): | |||
| y_xp = xp.asarray(y, dtype=x.dtype) | |||
| y_xp = xp.asarray(y, dtype=x.dtype, device=_compat.device(x)) | |||
There was a problem hiding this comment.
Untested fix, which only benefits eager JAX.
On jax.jit, device propagation fails due to jax-ml/jax#26000
In a follow-up PR I'll rework the test_device tests to align them to the pattern recently established in scipy.
| DASK = "dask.array", _compat.is_dask_namespace | ||
| SPARSE = "sparse", _compat.is_pydata_sparse_namespace | ||
| JAX = "jax.numpy", _compat.is_jax_namespace | ||
| JAX_GPU = "jax.numpy:gpu", _compat.is_jax_namespace |
There was a problem hiding this comment.
As an aside, I profoundly dislike that this enum is used in production by the dispatch mechanism.
@lucascolley would it be OK if I make it test-only again and just use is_*_namespace in dispatch?
There was a problem hiding this comment.
I profoundly dislike that this enum is used in production by the dispatch mechanism.
what motivates your dislike?
There was a problem hiding this comment.
This enum contains a lot of "backends" that are just variant duplicates, which is an artifact of using this enum to parametrize the xp fixture. Which is a decent hack for pytest, but which doesn't make any sense in the context of the dispatch system.
There was a problem hiding this comment.
okay, I see the problem. Happy to go with whatever works best for you. FWIW I would like to keep using an enum in the dispatch mechanism, but fair enough if it is too messy to maintain the like mapping.
|
(heads up, merge conflicts on the lock file will keep cropping up while renovate is spinning up, feel free to ignore) |
|
the renovate traffic should have quietened down from now on |
There was a problem hiding this comment.
Thanks Guido, LGTM!
I'll merge once scientific-python/specs#380 is resolved...
|
@rgommers please could you disable requiring signed commits on this repo, when you have time? Unfortunately, the advice from scientific-python/specs#380 didn't play out in practice. |
|
sure, done |
tests-cudaanddev-cudapixi environments, run all tests first on CPU and then on GPU:xpx.at__setitem__with bool mask and dtype mismatch fails pytorch/pytorch#150017