[bugfix] Fix megatron lr_mult (#9524)

Author: Jintao-Huang

swift/megatron/trainers/base.py

@@ -446,6 +446,8 @@ def _get_param_groups(
                 param_group['max_lr'] = lr
                 param_group['min_lr'] = min_lr
             lr_mult = param_group.pop('lr_mult')
+            # Instead of using lr_mult to control the learning rate, we directly use max_lr/min_lr.
+            param_group['lr_mult'] = 1.
             param_group['max_lr'] *= lr_mult
             param_group['min_lr'] *= lr_mult
         return param_groups