Fix possible sync issues with fasttensors

Author: turboderp

exllamav2/fasttensors.py

@@ -189,6 +189,8 @@ def get_tensor(self,
                    out_dtype = None) -> torch.Tensor:
         global global_tensorcache
 
+        torch.cuda.synchronize()
+
         if self.tensor_remap and (not_fast or not self.fast):
             key = self.tensor_remap[key]
 
@@ -236,4 +238,6 @@ def get_tensor(self,
                 global_tensorcache = global_tensorcache[1:]
             global_tensorcache.append((cachekey, tensor))
 
+        torch.cuda.synchronize()
+
         return tensor