My specific comment is under the context of the TVM's unfied IR infra, to elaborate further.
The current TVM's new IR Infra incorporates three variants of functions in the same IRModule, tir::PrimFunc, relay::Function, and we will have additional ExternFunc for MLIR variants of that to represent an MLIR dialect functions. And use the same set of type system to represent the function signatures so that we can freely enable calls among the functions. To enable a hybrid comobo translation, here is the ultimate TVM ingestion path that we can envision: start from TF GraphDef - Step 1: GraphDef -> tvm::IRModule: translate as much as GraphDef as we can, including high-level control flows, the rest parts of are collapsed as several `tvm::ExternGraphDefFunc`. The fragments becomes relay calls into these tvm::ExternGraphDefFunc - Step 2: Lower the remaining GraphDef fragments(in tvm::ExternFunc) to HLO (as tvm::ExternHLOFunc) - Step 3: Translate the rest of the HLO functions to relay Step 1 ensures us to cover high-level important operators, as well as control flow constructs so that they can be natively converted without losing the information. Step 2-3 ensures more coverage through HLO. By combining the two, we will be able to get the best of both worlds. Generalizing a bit, because we can represent different variants of functions under the same `tvm::IRModule`, such kind of multi-stage lowering might not only be relevant to TF, but also other frameworks with multi-stage IRs. The conventional view was to only tap into one of them, but the new design allows us to tap into multiple of them and gives more advantages under the performance/coverage tradeoffs. --- [Visit Topic](https://discuss.tvm.ai/t/rfc-mlir-frontend/6473/5) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/f67c1600f7d7d2b205b14a1863c13c1166c204da03b85367df48209b34d37919).
