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
Today, users of the Polars GPU engine don't experience graceful performance degradation for GPU out-of-memory errors like users of cuDF pandas do -- instead, the query will fail. This is a pain point that's come up consistently in feedback from early users.
In parallel to investigating UVM, we should also investigate the feasibility of the current inspector-executor design supporting different types of CPU fallback when the GPU engine hits an out-of-memory error. E.g., we could rewind and execute the full query from scratch with the CPU engine or we could rewind only a portion of the query (which might require subgraph based hybrid execution).
The text was updated successfully, but these errors were encountered:
Today, users of the Polars GPU engine don't experience graceful performance degradation for GPU out-of-memory errors like users of cuDF pandas do -- instead, the query will fail. This is a pain point that's come up consistently in feedback from early users.
In parallel to investigating UVM, we should also investigate the feasibility of the current inspector-executor design supporting different types of CPU fallback when the GPU engine hits an out-of-memory error. E.g., we could rewind and execute the full query from scratch with the CPU engine or we could rewind only a portion of the query (which might require subgraph based hybrid execution).
The text was updated successfully, but these errors were encountered: