CPU/GPU interop with RandomForest #6175
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
hcho3
changed the title
hcho3
added
improvement
|
Does "latest treelite" mean we need #6212 to be merged for this to work? |
|
This PR actually contains duplicate commits from #6212. So yes, we need the latest Treelite to support the interop feature. |
| # treelite does an internal isinstance check to detect an sklearn | ||
| # RF, which proxymodule interferes with. We work around that | ||
| # temporarily here just for treelite internal check and | ||
| # restore the __class__ at the end of the method. |
There was a problem hiding this comment.
In a future version of Treelite, we should add an alternative version of treelite.sklearn.import_model that takes in a second argument model_type? The caller would provide a string to uniquely identify the model type, so that we can remove the internal isinstance check.
There was a problem hiding this comment.
Maybe, depends on whether we do other changes in cuML itself to potentially avoid this
|
/merge |
There was a problem hiding this comment.
Just realized I forgot to review this PR before it was merged, sorry about that! I just took a look at it. It looks like several workarounds were used, but all of them seemed necessary. Also, good testing. LGTM. Thanks for the great work @hcho3!
First version for CPU/GPU interop with RandomForest. Note. This feature requires latest Treelite.