Disable offloading attention like kernels to rocMLIR if there is no wmma/mfma for requested dtype

[umangyadav]

DOR (Definition of Ready)

Ready

Description

rocMLIR doesn't support attention like kernels. eg. GEG, Attention, CEG if there is no associated wmma/mfma instruction for that dtype on the hardware.

e.g. using Fp32 on navi3x for the attention would crash.

DOD (Definition of Done)

Attention like kernels are not being generated for unsupported dtypes.

[pfultz2]

Right now there isnt an easy way to do this. If we skip attention in fuse_mlir there will be left over group op that wont get lowered. We probably need to make a pass inline_group to inline the attention afterwards, but then we will need to run fuse_pointwise_reduce pass again. This will need a tweak so fuse_pointwise uses a different name to avoid name conflicts on the second run. Probably move the counter to the pass manager.

So these would be the set of subtasks to get this to work:

• Add a inline_group pass to inline group operators
• Tweak fuse_pointwise to avoid name conflicts
• Add another fuse_pointwise_reduce pass