Support importing MLIR file or stream into Catalyst

[sengthai]

Summary

Catalyst should provide a way to import an MLIR file or stream directly into its workflow. This would enable users to leverage MLIR interoperability by ingesting pre-existing IRs or those generated by other tools.

Motivation

• Enable advanced transformations and analysis before ingestion.
• Improve flexibility for researchers and advanced users.

Expected behavior

• Users can import an MLIR file or provide an MLIR stream (e.g., via stdin, etc.) into Catalyst.
• Need to validate if the MLIR instructions are supported in the backend.

import catalyst.load

program_mlir_str = "MLIR CONTENT"
program_mlir_path = "path to mlir"

qjit_cir_1 = catalyst.load(program_mlir_str)
qjit_cir_2 = catalyst.load(filepath=program_mlir_path)

# Execute the program
qjit_cir_1()
qjit_cir_2() 

[erick-xanadu]

@sengthai not exactly what you requested but take a look at this PR https://github.com/PennyLaneAI/catalyst/pull/1334 :)

[sengthai]

@sengthai not exactly what you requested but take a look at this PR #1334 :)

Nice! Thanks @erick-xanadu! The core functions are there! I think that approach is helpful for debugging and for users who try to execute their MLIR as well. We should take further action on your PR.

[dime10]

Besides Erick's approach, it's hard to implement what you suggest above @sengthai because we need the type signature of the program in order to correctly provide arguments and interpret results. At some point I think we were inspecting the IR and attempt to deduce the type signature automatically, but it wasn't without issues (at least with the llvmir I remember it did not work).