[FEA] Use RMM directly instead of adapters

rapidsai / cuml

cuML - RAPIDS Machine Learning Library

https://docs.rapids.ai/api/cuml/stable/

Apache License 2.0

4.08k stars 525 forks source link

[FEA] Use RMM directly instead of adapters #3145

Open divyegala opened 3 years ago

divyegala commented 3 years ago

Linking issue from RAFT: https://github.com/rapidsai/raft/issues/97

Other things to consider:

RMM also has adapters for thrust device vectors which is especially useful when trying to use the thrust API in cuml code and creates a nice abstraction as I have also found out that we use explicit kernels for quite a good number of tasks that could be a couple liners with thrust
While the discussion to use more thrust/cub APIs wherever possible is definitely a much larger one, my proposal to this upgrade is that we first start with using RMM directly

Tagging @teju85 @JohnZed @dantegd @cjnolet for thoughts

dantegd commented 3 years ago

why not put the other things to consider in the original issue as well to avoid split discussions?

divyegala commented 3 years ago

why not put the other things to consider in the original issue as well to avoid split discussions?

I figured these other things were specific only to cuML for now, but if not then I'll just append it to the original issue

github-actions[bot] commented 3 years ago

This issue has been marked rotten due to no recent activity in the past 90d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

github-actions[bot] commented 3 years ago

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.