[INFER] llama&qwen2 A8W8 support skip_scale

llm/predict/predictor.py

@@ -1241,6 +1241,7 @@ def create_predictor(
             config.quant_type = predictor_args.quant_type
             config.cachekv_int8_type = predictor_args.cachekv_int8_type
             config.use_fake_parameter = predictor_args.use_fake_parameter
+            config.top_k = predictor_args.top_k
             config.single_card_ptq = True
             if config.quantization_config.quant_type is not None:
                 predictor_args.quant_type = config.quantization_config.quant_type

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 = []