Update inference.py

inference.py

@@ -137,7 +137,7 @@ class InferenceRecipe:
         """Run a warmup pass."""
         try:
             frame_size = int(self.sample_rate / self.frame_rate)
-            chunk = torch.zeros(1, 1, frame_size, dtype=torch.float32, device=self.device)
+            chunk = torch.zeros(1, 1, frame_size, dtype=torch.float32)
             codes = self.mimi.encode(chunk)
             
             with torch.no_grad(), self.lm_gen.streaming(1), self.mimi.streaming(1):
@@ -145,7 +145,6 @@ class InferenceRecipe:
                 if tokens is not None:
                     _ = self.mimi.decode(tokens[:, 1:])
                     
-            torch.cuda.synchronize()
             logger.info("Warmup pass completed")
 
         except Exception as e: