[REVIEW]Add support and tests for cuML and XGBoost

[VibhuJawa]

Add support and tests for cuML and XGBoost

This PR addresses #309

TODO:

• Ensure efficient path for non distributed models and allow single GPU xgboost to succeed

   For more context around this see issue: rapidsai/cuml#4406 .
  
   We were previously training on client which is:
     a. Very inefficient and possibly problematic in multi-node clusters and heterogeneous setup.
     b. Training non distributed xgboost models on dask collections is not supported.
  

• Test for singe GPU cuml model

• Test for multi gpu cuml model

• Test for singe gpu xgboost model

• Test for multi gpu xgboost model

• Update GPU-CI environment with the right libraries
  See PR: https://github.com/rapidsai/dask-build-environment/pull/16/files

Follow Up work to enable predict with multi GPU cuML models:

• Triage single gpu cuml model failure
  See issue: [BUG]Predict fails with multi GPU cuML models #332

[codecov-commenter]

Codecov Report

Merging #330 (2195a06) into main (96524ee) will increase coverage by 0.01%.
The diff coverage is 88.23%.

@@            Coverage Diff             @@
##             main     #330      +/-   ##
==========================================
+ Coverage   95.69%   95.70%   +0.01%     
==========================================
  Files          65       65              
  Lines        2854     2863       +9     
  Branches      534      536       +2     
==========================================
+ Hits         2731     2740       +9     
+ Misses         75       74       -1     
- Partials       48       49       +1     

Impacted Files	Coverage Δ
dask_sql/physical/rel/custom/create_model.py	93.22% <88.23%> (-2.94%)	⬇️
dask_sql/physical/utils/sort.py	90.62% <0.00%> (+7.29%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 96524ee...2195a06. Read the comment docs.

[VibhuJawa]

CC: @charlesbluca , How do i go about adding xgboost and cuML to the gpu-CI we have setup ?

[charlesbluca]

You can open a PR adding these packages to the environment created in this dockerfile:

https://github.com/rapidsai/dask-build-environment/blob/main/dask_sql.Dockerfile

[VibhuJawa]

@charlesbluca , This is now ready for review.

This PR is dependent on https://github.com/rapidsai/dask-build-environment/pull/16/files

[VibhuJawa]

For more context around this see issue: rapidsai/cuml#4406 .

We were previously training on client which is:

a. Very inefficient and possibly problematic in multi-node clusters and heterogeneous setup.
b. Training non distributed xgboost models on dask collections is not supported.

[charlesbluca]

Thanks for the work here @VibhuJawa 😄 a couple comments, I'm not the most knowledgeable on Dask ML stuff so would feel good getting a second review on this:

[ChrisJar]

LGTM!

[charlesbluca]

Missed this. but generally things look good here! Happy to merge this once cuML / XGBoost are successfully added to gpuCI and tests are passing

[VibhuJawa]

@GPUtester rerun tests .

[charlesbluca]

Think we need to mark these as fixtures to expose them to tests?

[charlesbluca]

rerun tests

[charlesbluca]

rerun tests