Reactant can already go directly to a tensorflow saved model!: Serialization | Reactant.jl
And yeah essentially all we need is a stablehlo for whatever foreign code is desired to be called. So if we can get that from the pytorch code, then it can just be hooked in via Reactant’s hlo_call FFI. See Reactant.jl/ext/ReactantPythonCallExt/pycall.jl at 13771b008e590945717f986b9e55ab2f466959a6 · EnzymeAD/Reactant.jl · GitHub for how its used for jax interop.