Closed Groverkss closed 1 week ago
This patch generalizes the AMDGPUChainedMatmul pass to use VectorContractOpInfo to query and transpose dims, instead of hardcoding indexing maps.
Since this introduces a n, m matmul, I need to check if attention still works with this patch.
n, m
Actually, CI seems to be passing on this, so I'm going to land this
This patch generalizes the AMDGPUChainedMatmul pass to use VectorContractOpInfo to query and transpose dims, instead of hardcoding indexing maps.