Merge branch 'jbarker/debug_pp_convert' into 'main'

Fix bug when loading pp>1 model with frozen layers See merge request ADLR/megatron-lm!2523
NVIDIA · Jan 9, 2025 · 3046e33 · 3046e33
2 parents 8fba594 + 458bfc9
commit 3046e33
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/megatron/core/optimizer/optimizer.py b/megatron/core/optimizer/optimizer.py
@@ -374,7 +374,8 @@ def get_loss_scale(self):
         return self.grad_scaler.scale
 
     def reload_model_params(self):
-        self._copy_model_params_to_main_params()
+        if self.param_groups:
+            self._copy_model_params_to_main_params()
 
     def _unscale_main_grads_and_check_for_nan(self):