fix: skip when gradient is sparse

Author: kozistr

pytorch_optimizer/optimizer/experimental/ranger25.py

@@ -103,7 +103,7 @@ def schedule_beta3(t_alpha_beta3: Optional[float], step: int, beta1: float, beta
     @torch.no_grad()
     def orthogonalize_gradients(self, params, eps: float = 1e-16) -> None:
         for p in params:
-            if p.grad is None:
+            if p.grad is None or p.grad.is_sparse:
                 continue
 
             w = p.view(-1)

pytorch_optimizer/optimizer/orthograd.py

@@ -52,7 +52,7 @@ def reset(self):
     @torch.no_grad()
     def orthogonalize_gradients(self, params) -> None:
         for p in params:
-            if p.grad is None:
+            if p.grad is None or p.grad.is_sparse:
                 continue
 
             w = p.view(-1)