Merge branch 'finetune_fix' into 'main'

Update code used for finetuning to latest API.

See merge request ADLR/megatron-lm!174

megatron/model/classification.py

@@ -52,8 +52,7 @@ class Classification(MegatronModule):
 
     def forward(self, input_ids, attention_mask, tokentype_ids):
 
-        extended_attention_mask = bert_extended_attention_mask(
-            attention_mask, next(self.language_model.parameters()).dtype)
+        extended_attention_mask = bert_extended_attention_mask(attention_mask)
         position_ids = bert_position_ids(input_ids)
 
         _, pooled_output = self.language_model(input_ids,

megatron/model/multiple_choice.py

@@ -64,8 +64,7 @@ class MultipleChoice(MegatronModule):
         attention_mask = attention_mask.view(-1, attention_mask.size(-1))
         tokentype_ids = tokentype_ids.view(-1, tokentype_ids.size(-1))
 
-        extended_attention_mask = bert_extended_attention_mask(
-            attention_mask, next(self.language_model.parameters()).dtype)
+        extended_attention_mask = bert_extended_attention_mask(attention_mask)
         position_ids = bert_position_ids(input_ids)
 
         _, pooled_output = self.language_model(input_ids,

tasks/finetune_utils.py

@@ -161,15 +161,15 @@ def _train(model, optimizer, lr_scheduler, forward_step,
             start_iteration = 0
 
             # Train for one step.
-            losses_dict, _ = train_step(forward_step, batch, model,
-                                        optimizer, lr_scheduler)
+            losses_dict, skipped_iter = train_step(forward_step, batch, model,
+                                                   optimizer, lr_scheduler)
             iteration += 1
 
             # Logging.
             report_memory_flag = training_log(losses_dict, losses_dict_sum,
                                               optimizer.param_groups[0]['lr'],
                                               iteration, optimizer.loss_scale,
-                                              report_memory_flag)
+                                              report_memory_flag, skipped_iter)
 
             # Autoresume
             if args.adlr_autoresume and \