[RFC] XLA:GPU runtime support for CUDA Graphs and LMHLO removal

[ezhulenev]

Link: XLA:GPU Runtime 2023-2024

Hi everyone,

As a part of modernizing and simplifying XLA:GPU runtime we'll be working on a project to bring CommandBuffers support to XLA (mapped to CUDA graphs for NVIDIA backend).

Also as a part of this project we want to deprecate and remove intermediate LMHLO step from the XLA:GPU compilation pipeline:

1. As we do not plan to use MLIR to lower from HLO to Thunks and CommandBuffers this step is not required
2. We believe that we'll simplify the overall stack by removing non-load-bearing MLIR dialect from the compilation path
3. As LMHLO is one of the first dialects designed with MLIR (while MLIR was very young), it's not the best example of building a dialect

Please let us know what you think and what concerns do you have.