MLIR plugin with XLA


I just find that the commit dropped the implementation of the mlir_gpu plugin prototype in XLA.

Could I ask why you are considering drop the code and is there any other plan for XLA runtime to use MLIR optimization? @herhut

Copying what @herhut wrote on another mailing list:

The changes you have seen landing are only removing the integration glue code between XLA and MLIR that was written for the prototype. This integration was incomplete and unmaintained. The actual substance of the prototype, which are the lowerings, the dialects that were used, etc., were already moved to different places before this change.

1 Like