Fix: backward() inside @torch .no_grad() — use torch.enable_grad() for dense gradient computation 39301d6 verified theapemachine commited on May 1