switch back

config.json

@@ -69,9 +69,9 @@
   "num_attention_heads": 12,
   "num_hidden_layers": 22,
   "num_initial_layers": 1,
-  "pad_logits": false,
+  "pad_logits": true,
   "pad_token_id": 0,
-  "padding": "padded",
+  "padding": "unpadded",
   "pooling_type": "cls",
   "position_embedding_type": "absolute",
   "rotary_emb_base": 10000.0,
@@ -82,7 +82,7 @@
   "sliding_window": 128,
   "transformers_version": "4.44.1",
   "type_vocab_size": 2,
-  "unpad_embeddings": false,
+  "unpad_embeddings": true,
   "use_cache": true,
   "use_fa2": true,
   "use_sdpa_attn_mask": false,

modeling_flexbert.py

@@ -1650,6 +1650,7 @@ class FlexBertForCausalLM(FlexBertPreTrainedModel):
                 input_ids, attention_mask, position_ids, labels
             )
 
+        
         hidden_states = self.bert(
             input_ids,
             attention_mask=None,