Currently, TT-MLIR supports extracting const-eval-able operation subgraphs into functions whose results are being cached and reused across forward invocations.
Additionally, TT-MLIR supports hoisting execution of certain operations to the CPU on the host machine, which is currently being used to execute some non-supported SHLO ops through SHLO -> linalg -> ... -> LLVM path.
This issue encapsulates the efforts required for enabling CPU-hoisting of the const-eval subgraphs.