Default behavior for sparseml w/o recipes (#13)

Author: natuan

src/transformers/sparse.py

@@ -111,17 +111,15 @@ def create_scheduler(self, num_training_steps: int):
             # scheduler already set
             return
 
-        if self.manager.learning_rate_modifiers:
+        if self.manager is not None and self.manager.learning_rate_modifiers:
             # allow SparseML to manage LR and set a dummy scheduler
-            self.lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
-                self.optimizer, lambda _: 1.0, -1
-            )
+            self.lr_scheduler = torch.optim.lr_scheduler.LambdaLR(self.optimizer, lambda _: 1.0, -1)
         else:
             # default scheduler
             super().create_scheduler(num_training_steps)
 
     def qat_active(self, epoch: int):
-        if not self.manager.quantization_modifiers:
+        if self.manager is None or not self.manager.quantization_modifiers:
             return False
 
         qat_start = min([mod.start_epoch for mod in self.manager.quantization_modifiers])