diff --git a/server/text_generation_server/utils/gptq/exllamav2.py b/server/text_generation_server/utils/gptq/exllamav2.py
index 1e0be490..dc8353f3 100644
--- a/server/text_generation_server/utils/gptq/exllamav2.py
+++ b/server/text_generation_server/utils/gptq/exllamav2.py
@@ -143,7 +143,7 @@ class QuantLinear(nn.Module):
         #     self.bias = None
 
     # def post_init(self, temp_dq):
-        temp_dq = ExLlamaV2DeviceTensors(self.qweight.device.index , self.temp_dq_size())
+        temp_dq = ExLlamaV2DeviceTensors(self.qweight.device.index , self.temp_dq_size() + self.temp_fwd_size(4096, 8))
         assert self.qweight.device.type == "cuda"
         assert self.qweight.device.index is not None
         self.q_tensors = {
@@ -152,7 +152,7 @@ class QuantLinear(nn.Module):
             "scales":self.scales,
             "g_idx":self.g_idx
         }
-        temp_dq = temp_dq.get_scratch_slice(self.temp_dq_size())
+        temp_dq = temp_dq.get_scratch_slice(self.temp_dq_size() + self.temp_fwd_size(4096, 8))
         self.q_handle = ext_make_q_matrix(
             self.q_tensors, temp_dq
         )