Hello all,
I would like to know what to expect regarding using Mooncake for reverse AD on GPUs, preferably via KernelAbstractions. I see there is a CUDA extension handling CuArrays. What works now, and what could work in the next few months ?
Thanks for the great work on Mooncake / KernelAbstractions.
Hello, I may be wrong but, unlike Enzyme, Mooncake can’t diff through the llvm / ptx code done by GPU backends, the only possibility would be to add a thin Enzyme layer to Mooncake but may be better off doing that in DI directly (use Enzyme to buld_rule for any backend).
This wouldn’t be that much of an issue if the rule system was easy enough and I think that’s what should be worked on rn.