Fixing bug in Megatron BERT when loss mask is all zeros

nemo/collections/nlp/data/language_modeling/megatron/dataset_utils.py

@@ -234,6 +234,10 @@ def create_masked_lm_predictions(
         return (output_tokens, masked_lm_positions, masked_lm_labels, token_boundary)
 
     num_to_predict = min(max_predictions_per_seq, max(1, int(round(len(tokens) * masked_lm_prob))))
+    if num_to_predict < 1:
+        logging.info(
+            F' > WARNING: number of tokens is : {len(tokens)} and mask_probability is {masked_lm_prob}. None of the tokens will be masked'
+        )
 
     ngrams = np.arange(1, max_ngram_size + 1, dtype=np.int64)
     if not geometric_dist:

nemo/collections/nlp/models/language_modeling/megatron_bert_model.py

@@ -358,7 +358,13 @@ def loss_func(self, loss_mask, sentence_order, output_tensor):
 
         lm_loss_ = lm_loss_.float()
         loss_mask = loss_mask.float()
-        lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
+
+        # Sometimes when the number of tokens is very small, none of the tokens get masked for prediction. In that case loss mask is all zeros
+        # i.e Happens when the entire batch is masked out (Practically when MBS=1 or 2, and the number of tokens in each batch is < 7 )
+        if loss_mask.sum() == 0:
+            lm_loss = torch.sum(lm_loss_.view(-1)) * 0.0
+        else:
+            lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
 
         if sop_logits is not None:
             sop_loss = F.cross_entropy(sop_logits.view(-1, 2).float(), sentence_order.view(-1), ignore_index=-1)