update

ppfleetx/models/language_model/gpt/dygraph/hybrid_model.py

@@ -907,6 +907,10 @@ def __init__(self,
 
         if no_recompute_layers is None:
             no_recompute_layers = []
+        else:
+            if recompute_granularity == 'full':
+                assert len(no_recompute_layers) == 0, \
+                    "for pp with full recompute, no_recompute_layers is not support"
 
         assert sequence_parallel is False, "Sequence parallel strategy \
                     is not supported in GPTForPretrainingPipe model now."
@@ -978,7 +982,6 @@ def _logits_helper(embedding, output):
                 "mp_group": fleet.fleet._hcg.get_model_parallel_group(),
                 "offload": False,
                 "partition": False,
-                "no_recompute_layers": no_recompute_layers
             },
             num_virtual_pipeline_stages=virtual_pp_degree)