update fp16 model layout conversion conditions

vllm/model_executor/models/baichuan.py

@@ -413,7 +413,7 @@ class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA):
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
                 
-        if self.use_llama_nn:
+        if self.use_llama_nn and self.quant_method is None :
             lay_key_words = [
                 "self_attn.W_pack.weight",
                 "self_attn.o_proj.weight",

vllm/model_executor/models/chatglm.py

@@ -411,7 +411,7 @@ class ChatGLMForCausalLM(nn.Module, SupportsLoRA):
                                     default_weight_loader)
             weight_loader(param, loaded_weight)
         
-        if self.use_llama_nn:
+        if self.use_llama_nn and self.quant_method is None:
             lay_key_words = [
                 "self_attention.query_key_value.weight",
                 "self_attention.dense.weight",

vllm/model_executor/models/llama.py

@@ -531,7 +531,7 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA):
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)  
             
-        if self.use_llama_nn:
+        if self.use_llama_nn and self.quant_method is None :
             lay_key_words = [
                 "self_attn.qkv_proj.weight",
                 "self_attn.o_proj.weight",

vllm/model_executor/models/qwen.py

@@ -355,7 +355,7 @@ class QWenLMHeadModel(nn.Module):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
-        if self.use_llama_nn:
+        if self.use_llama_nn and self.quant_method is None :
             lay_key_words = [
                 "attn.c_attn.weight",
                 "attn.c_proj.weight",

vllm/model_executor/models/qwen2.py

@@ -448,7 +448,7 @@ class Qwen2ForCausalLM(nn.Module, SupportsLoRA):
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
                 
-        if self.use_llama_nn:
+        if self.use_llama_nn and self.quant_method is None:
             lay_key_words = [
                 "self_attn.qkv_proj.weight",
                 "self_attn.o_proj.weight",