Paul Fultz II
Paul Fultz II
> since other matcher can depend on that specific structure, so they will need to be extended as well probably. This same rewriting is needed for layernorm, Right now we...
> Also, this math rewrite only a partial solution, since larger numbers still can cause problem. I would suggest having a dedicated rmsnorm fp16-to-fp32 convert optional pass for those cases....
@kahmed10 Any feedback on this?
We probably need to add this operator to mlir, but it still is probably good to check for supported pointwise before fusing attention.
Its a change that needs to happen in mlir, but it should be simple to add.
The goal of this is to fuse layernorm with with two convs/gemms. In #3010, it will fuse whats after the reductions with the following convolution, but we still need to...
> 1. Fuse reshapes first. convolution + reshapes + pointwise --> Always enable this, use https://github.com/ROCm/AMDMIGraphX/pull/3010 to use fuse module functionality. This address issue #2822.
Related #3844, #3920
We use the order in the onnx files to ensure that the order is consistent with the onnx file: https://github.com/ROCm/AMDMIGraphX/pull/479 For invalid onnx file, we can write a python script...
What is the compiler error? I dont see any error.