You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
. I believe what's happening is that this is an import in the global level where dask-cuda can't ensure the CUDA context has already been created, in turn leaving that task then to cuDF.
@madsbk could you check if it's possible to move that import from the global level to a local one, in such a way as to prevent cudf from ever being imported before dask-cuda has had a chance to initialize the CUDA context? I took a quick look but it seems that this is an indirect dependency and I wasn't sure how to fix it without risking breaking something else.
The text was updated successfully, but these errors were encountered:
Fixes#491 by moving dispatch of cudf objects within a `register_lazy("cudf")`.
Authors:
- Mads R. B. Kristensen <[email protected]>
Approvers:
- Peter Andreas Entschev (@pentschev)
URL: #492
Latest nightly tests failed due to CUDA initialization error, which seems to be rooted at
dask-cuda/dask_cuda/proxify_device_objects.py
Line 92 in cd8a87d
@madsbk could you check if it's possible to move that import from the global level to a local one, in such a way as to prevent
cudf
from ever being imported before dask-cuda has had a chance to initialize the CUDA context? I took a quick look but it seems that this is an indirect dependency and I wasn't sure how to fix it without risking breaking something else.The text was updated successfully, but these errors were encountered: