🐛 Bug

ThunderFX fallback submodules are NOT being compiled by Inductor and apparently executed in eager mode instead.

This is causing missed opportunities of fusion (such as @kshitij12345 found in #2527 (comment)) and memory spikes (#2501).

The impact will be huge when combined with activation checkpointing. In the current implementation, if Thunder cannot support all ops in a checkpointed region, the entire region falls back to torch.compile. When checkpointing is applied in the granularity of blocks, a huge part of a model will not be compiled by neither Thunder nor Inductor.

cc @kshitij12345 @IvanYashchuk