add recompute of postnorm in pp (#10849)

Author: liuruyan

paddlenlp/transformers/deepseek_v2/modeling.py

@@ -2040,7 +2040,9 @@ def self_attn_compute(self, hidden_states, **kwargs):
         hidden_states = residual + hidden_states
 
         residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
+
+        if not self.using_post_norm_recompute:
+            hidden_states = self.post_attention_layernorm(hidden_states)
 
         return hidden_states, residual