JIT LTO Exploratory Work

[divyegala]

Using @robertmaynard's work in https://github.com/robertmaynard/runtime_lto_examples, we need to work on one or two examples in the cuML repository and transforming the usage to work seamlessly with a JIT LTO module. Once we have a working algorithm e2e, we need to measure:

• Binary size
• Compilation time
• Runtime performance (benchmarks, HPO, etc.)

Questions to answer in the future:

• How will we ship intermediate fatbins
• Can we cache JIT to disk