In some cases it seems cudaPointerGetAttributes is returning cudaSuccess and not setting the last error, which means if something else has previously caused an error with cuda, you blow up here.

I don't really know what the cause of that previous error is, but the situation is this:

• on a machine using VMware horizon (which exposes a vgpu that is cuda enabled, but cuda not working)
• using xgboost compiled for cuda, because when running on grid compute we want to use cuda
• running some unit tests that attempt to use xgboost on CPU
• attempt to construct DMatrix object with normal numpy array (in CPU memory)
• blows up on this check, completely preventing use of xgboost

I suppose alternatively you could call cudaGetLastError before cudaPointerGetAttributes in addition to the existing check.

I suppose alternatively you could call cudaGetLastError before cudaPointerGetAttributes in addition to the existing check.

This sounds like a better workaround, would you like to help make the change?

@hcho3 can this be cherry-picked into the next 3.x release please?

I created #11494 to keep track of the upcoming patch release. Until the new patch release is out, consider using a nightly distribution instead: https://xgboost.readthedocs.io/en/release_3.0.0/install.html#id1