增加lm_nn的量化控控控制,控制为tn

vllm/model_executor/models/gpt_neox.py

@@ -360,6 +360,7 @@ class GPTNeoXForCausalLM(nn.Module, SupportsPP):
             loaded_params.add(name)
         #当为triton支持推理的时候不能进行处理
         if self.quant_method == "compressed_tensors":
+            os.environ['LM_NN'] = '0'
             lay_key_words = [
                 "attention.query_key_value.weight",
                 "attention.dense.weight",

vllm/model_executor/models/llama.py

@@ -550,6 +550,7 @@ class LlamaModel(nn.Module):
             
         #当为triton支持推理的时候不能进行处理
         if self.quant_method == "compressed_tensors":
+            os.environ['LM_NN'] = '0'
             lay_key_words = [
                 "self_attn.qkv_proj.weight",
                 "self_attn.o_proj.weight",

vllm/model_executor/models/qwen.py

@@ -1174,6 +1174,7 @@ class QWenBaseModel(nn.Module, SupportsPP, SupportsLoRA):
                         qweight.data=torch.cat((qweight.data,qweight_pad),dim=1).contiguous()
                        
         if self.quant_method == "compressed_tensors":
+            os.environ['LM_NN'] = '0'
             lay_key_words = [
                 "attn.c_attn.weight",
                 "attn.c_proj.weight",

vllm/model_executor/models/qwen2.py

@@ -527,6 +527,7 @@ class Qwen2Model(nn.Module):
                         qweight.data=torch.cat((qweight.data,qweight_pad),dim=1).contiguous()
                    
         if self.quant_method == "compressed_tensors":
+            os.environ['LM_NN'] = '0'
             lay_key_words = [
                 "self_attn.qkv_proj.weight",
                 "self_attn.o_proj.weight",