This is likely because you need to your configure optimizers from
def configure_optimizers(self):
return YourOptimizer(self.parameters(), ...)
to
def configure_optimizers(self):
return YourOptimizer(self.trainer.model.parameters(), ...)
This is a quirk that is currently necessary in Lightning, but as FSDP matures in the future, this won’t be necessary anymore.
If this doesn’t solve your issue, please feel free to post a GitHub bug report and we can take a closer look.