Breanna Devore-McDonald

Flash decoding round 1

3

First round in addressing #4334 ## Motivation rocMLIR has implemented split kv and GQA, which enables us to implement flash decoding. Now we need to add this to the migraphx...

G+G fusion in rocMLIR supports optional elementwise fusion between gemms, and before/after either gemm. The current MIGraphX implementation has the elementwise op between gemms as required

Currently, g+g fusion is enabled by a user switch. We need to change this to automatically decide when to apply this fusion based on what the nodes look like; i.e....

Current version of protobuf has a security vulnerability so need to upgrade to >22