[INFER] llama&qwen2 A8W8 support skip_scale

llm/predict/predictor.py

@@ -886,12 +886,8 @@ def init_model_inputs(self, config: PredictorArgument):
                 ]
             )
             # self.model_inputs["src_mask/tgt_mask"] is read only, will not be updated!
-            src_mask = (
-                alibi_encoder + (1 - src_mask) * paddle.finfo(self.dtype).min
-            ).cast(self.dtype)
-            tgt_mask = (
-                alibi_decoder + (1 - tgt_mask) * paddle.finfo(self.dtype).min
-            ).cast(self.dtype)
+            src_mask = (alibi_encoder + (1 - src_mask) * paddle.finfo(self.dtype).min).cast(self.dtype)
+            tgt_mask = (alibi_decoder + (1 - tgt_mask) * paddle.finfo(self.dtype).min).cast(self.dtype)
             self.model_inputs["rope_emb"] = paddle.concat([src_mask.reshape([-1]), tgt_mask.reshape([-1])])
 
     def _preprocess(self, input_text: list[str]):

paddlenlp/experimental/transformers/bloom/modeling.py

@@ -293,6 +293,7 @@
 
     @paddle.no_grad()
     def set_state_dict(self, state_dict, use_structured_name=True):
+        self.transformer_block.init_weight()
         for k, v in state_dict.items():
             if k.find("word_embeddings.weight") >= 0:
                 self.word_embeddings.weight.set_value(paddle.to_tensor(v))

paddlenlp/experimental/transformers/chatglm/modeling.py

@@ -377,6 +377,7 @@
 
     @paddle.no_grad()
     def set_state_dict(self, state_dict, use_structured_name=True):
+        self.transformer_block.init_weight()
         dtype = paddle.get_default_dtype()
         config = self.config
         embed_dim = config.hidden_size

paddlenlp/experimental/transformers/chatglm_v2/modeling.py

@@ -290,6 +290,8 @@
 
     @paddle.no_grad()
     def set_state_dict(self, state_dict):
+        self.transformer_block.init_weight()
+
         # find the real name.
         def key(name):
             result_list = []