fix galore layerwise with frozen params (#29743)

src/transformers/optimization.py

@@ -385,6 +385,7 @@ def get_scheduler(
                 scheduler_dict[param].step()
 
         for param in optimizer_dict.keys():
+            if param.requires_grad:
                 param.register_post_accumulate_grad_hook(scheduler_hook)
 
         return LayerWiseDummyScheduler()

src/transformers/trainer.py

@@ -1303,6 +1303,7 @@ class Trainer:
                         optimizer_dict[param].zero_grad()
 
                 for param in model.parameters():
+                    if param.requires_grad:
                         param.register_post_accumulate_grad_hook(optimizer_hook)
 
                 optimizer_cls = LayerWiseDummyOptimizer