This change has caused issues especially with LoRAs that have dynamic rank. Will now be disabled by default, to allow full graph with unmerged LoRAs the option to allow compile is available in the Torch Compile Settings -node
Fp8 matmul (fp8_fast) doesn't seem feasible with unmerged LoRAs as you'd need to first upcast, then apply LoRA, then downcast back to fp8 and that is too slow. Direct adding in fp8 is also not possible since that's just not something fp8 dtypes support.
Not really getting great results yet with MagCache, but at least pretty much on bar with TeaCache so it can be an option that hopefully improves in time
kijai