Re^3: Do you have AI::MXNet running?

[bliako]

a real pain!

the signature for MXOptimizeForBackend in c_api.h differs wildly from the parameters passed in mxnet_wrap.cxx

I have solved that by updating the signature in mxnet.i to what c_api.h has. But I still don't know until the other dependency AI::NNVMCAPI compiles. If that hack does not solve it then we can try wrap the new signature with the old one.

Right now, AI::NNVMCAPI requires tvm, as you mentioned. My last attempt is getting tvm from the git repository (https://github.com/apache/tvm), install Intel mkl, dnnl etc. and see. There are problems right now with tvm compilation as I have encountered this: https://github.com/apache/tvm/issues/16556 because I have set DNNL ON.

All in all, it seems Apache have created millions of lines of code which are so smelly with bitrot is unbearable. Is the attic the right place or the cold ground instead?

[The_Dj]

In case others find there way here and we don't get to the bottom of this:

Apparently Nvidia changed MXOptimizeForBackend with the switch to cuda 12
and that's what's broken AI::MXNetCAPI

There is a beta of mxnet 2.0 available at https://archive.apache.org/dist/incubator/mxnet/2.0.0.beta1/
But it has its own compile issues... and it's a beta, so...

HTH someone

[bliako]

It's the end of the road for me for now. tvm does not compile for me complaining mostly about intel's oneapi api changes in its latest version (and intel's oneapi older versions are only available if you pay). On top of that the idiotic decision to use cmake makes it cumbersome to try different CFLAGS/LDFLAGS settings. If you get lucky with tvm please post your settings.