Update run_translation_no_trainer.py (#18637)

* Update run_translation_no_trainer.py

found an error in selecting `no_decay` parameters and some small modifications when the user continues to train from a checkpoint

* fixs `no_decay` and `resume_step` issue

1. change `no_decay` list
2. if use continue to train their model from provided checkpoint, the `resume_step` will not be initialized properly if `args.gradient_accumulation_steps != 1`

examples/pytorch/language-modeling/run_clm_no_trainer.py

@@ -464,7 +464,7 @@ def main():
 
     # Optimizer
     # Split weights in two groups, one with weight decay and the other not.
-    no_decay = ["bias", "LayerNorm.weight"]
+    no_decay = ["bias", "layer_norm.weight"]
     optimizer_grouped_parameters = [
         {
             "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
@@ -558,10 +558,15 @@ def main():
             starting_epoch = int(training_difference.replace("epoch_", "")) + 1
             resume_step = None
         else:
-            resume_step = int(training_difference.replace("step_", ""))
+            # need to multiply `gradient_accumulation_steps` to reflect real steps
+            resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
             starting_epoch = resume_step // len(train_dataloader)
             resume_step -= starting_epoch * len(train_dataloader)
 
+    # update the progress_bar if load from checkpoint
+    progress_bar.update(starting_epoch * num_update_steps_per_epoch)
+    completed_steps = starting_epoch * num_update_steps_per_epoch
+
     for epoch in range(starting_epoch, args.num_train_epochs):
         model.train()
         if args.with_tracking:
@@ -570,7 +575,9 @@ def main():
             # We need to skip steps until we reach the resumed step
             if args.resume_from_checkpoint and epoch == starting_epoch:
                 if resume_step is not None and step < resume_step:
-                    completed_steps += 1
+                    if step % args.gradient_accumulation_steps == 0:
+                        progress_bar.update(1)
+                        completed_steps += 1
                     continue
 
             with accelerator.accumulate(model):

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -602,10 +602,15 @@ def main():
             starting_epoch = int(training_difference.replace("epoch_", "")) + 1
             resume_step = None
         else:
-            resume_step = int(training_difference.replace("step_", ""))
+            # need to multiply `gradient_accumulation_steps` to reflect real steps
+            resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
             starting_epoch = resume_step // len(train_dataloader)
             resume_step -= starting_epoch * len(train_dataloader)
 
+    # update the progress_bar if load from checkpoint
+    progress_bar.update(starting_epoch * num_update_steps_per_epoch)
+    completed_steps = starting_epoch * num_update_steps_per_epoch
+
     for epoch in range(starting_epoch, args.num_train_epochs):
         model.train()
         if args.with_tracking:
@@ -614,7 +619,9 @@ def main():
             # We need to skip steps until we reach the resumed step
             if args.resume_from_checkpoint and epoch == starting_epoch:
                 if resume_step is not None and step < resume_step:
-                    completed_steps += 1
+                    if step % args.gradient_accumulation_steps == 0:
+                        progress_bar.update(1)
+                        completed_steps += 1
                     continue
 
             with accelerator.accumulate(model):

examples/pytorch/translation/run_translation_no_trainer.py

@@ -510,7 +510,7 @@ def main():
 
     # Optimizer
     # Split weights in two groups, one with weight decay and the other not.
-    no_decay = ["bias", "LayerNorm.weight"]
+    no_decay = ["bias", "LayerNorm.weight", "layer_norm.weight"]
     optimizer_grouped_parameters = [
         {
             "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
@@ -607,10 +607,15 @@ def main():
             starting_epoch = int(training_difference.replace("epoch_", "")) + 1
             resume_step = None
         else:
-            resume_step = int(training_difference.replace("step_", ""))
+            # need to multiply `gradient_accumulation_steps` to reflect real steps
+            resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
             starting_epoch = resume_step // len(train_dataloader)
             resume_step -= starting_epoch * len(train_dataloader)
 
+    # update the progress_bar if load from checkpoint
+    progress_bar.update(starting_epoch * num_update_steps_per_epoch)
+    completed_steps = starting_epoch * num_update_steps_per_epoch
+
     for epoch in range(starting_epoch, args.num_train_epochs):
         model.train()
         if args.with_tracking:
@@ -619,7 +624,9 @@ def main():
             # We need to skip steps until we reach the resumed step
             if args.resume_from_checkpoint and epoch == starting_epoch:
                 if resume_step is not None and step < resume_step:
-                    completed_steps += 1
+                    if step % args.gradient_accumulation_steps == 0:
+                        progress_bar.update(1)
+                        completed_steps += 1
                     continue
             outputs = model(**batch)
             loss = outputs.loss