VeRA adapters fail with a RuntimeError due to dtype mismatch when the base model is loaded in torch.float16…

VeRA adapters fail with a RuntimeError due to dtype mismatch when the base model is loaded in torch.float16…: a task in LegoFlow-SWE (Harbor dataset). Here sliced A and sliced B originate from the shared weight buffers (e.g., vera A , vera B ) and remain in the original half‑precision dtype, while…

The task

Here `sliced_A` and `sliced_B` originate from the shared weight buffers (e.g., `vera_A`, `vera_B`) and remain in the original half‑precision dtype, while other tensors in the expression get cast to `float32`. The auto‑casting logic that normally promotes adapter weights to float32 does not cover these buffers because…

Part of Lego-X/LegoFlow-SWE.