[GPT-3/MOE] Adapt recompute for latest paddle (#3191)

Author: sljlp

examples/language_model/gpt-3/dygraph/modeling.py

@@ -1178,6 +1178,4 @@ def _logits_helper(embedding, output):
                          loss_fn=GPTPretrainingCriterionPipe(),
                          topology=topology,
                          seg_method="layer:TransformerDecoderLayer",
-                         recompute_interval=1 if use_recompute else 0,
-                         recompute_partition=False,
-                         recompute_offload=False)
+                         recompute_interval=1 if use_recompute else 0)

examples/language_model/gpt-3/dygraph/run.sh

@@ -23,5 +23,5 @@ python -m paddle.distributed.launch --log_dir $log_dir --gpus "0,1,2,3,4,5,6,7"
     --sharding_degree 1\
     --use_pure_fp16 True\
     --use_recompute False\
-    --sharding_stage 2\
+    --sharding_stage 1\
     --sharding_offload False

examples/language_model/moe/dygraph/modeling.py

@@ -456,7 +456,7 @@ def __init__(self,
             }
             self.moe_mlp = MoeLayer(d_model=d_model,
                                     experts=experts_list,
-                                    gate_config=gate_config,
+                                    gate=gate_config,
                                     moe_group=moe_group,
                                     mp_group=mp_group,
                                     recompute_interval=self.recompute_interval)