修改awq 申请sharememshapce方式

vllm/model_executor/layers/quantization/awq.py

@@ -9,10 +9,22 @@ from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.utils import set_weight_attrs
 
+class AWQShareWorkSpace:
+    _instance = None
+    
+    def __new__(cls, *args, **kwargs):
+        if cls._instance is None:
+            cls._instance = super(AWQShareWorkSpace, cls).__new__(cls, *args, **kwargs)
+            # 执行初始化
+            cls._instance._initialize()
+        return cls._instance
+
+    def _initialize(self):
+        self.awqworkshapcesize = 2 << 29
+        self.awqworkshapce = torch.zeros(self.awqworkshapcesize // 2 + 1, dtype=torch.float16).cuda()
+        #print("AWQShareWorkSpace _initialize\n")
+        #print("self.awqworkshapce.device:",self.awqworkshapce.device)
 
-class AWQShareWorkSpace():
-    awqworkshapcesize=2<<29  #
-    awqworkshapce=torch.zeros(awqworkshapcesize//2+1,dtype=torch.float16).cuda()
 
 
 class AWQConfig(QuantizationConfig):
@@ -86,6 +98,7 @@ class AWQLinearMethod(LinearMethodBase):
 
     def __init__(self, quant_config: AWQConfig):
         self.quant_config = quant_config
+        self.awqsingleton= AWQShareWorkSpace()
 
     def create_weights(self, layer: torch.nn.Module,
                        input_size_per_partition: int,
@@ -197,8 +210,8 @@ class AWQLinearMethod(LinearMethodBase):
                             k,
                             self.quant_config.group_size,
                             padding_group,
-                            AWQShareWorkSpace.awqworkshapce,
-                            AWQShareWorkSpace.awqworkshapcesize)
+                            self.awqsingleton.awqworkshapce,
+                            self.awqsingleton.awqworkshapcesize)
         #下面是采用rocblas的做法
         # deqweight=ops.dequant_w4_gemm_colmajor(    #shape[n,k/8]--->[n,k]
         #                   qweight,