clean resources properly

generate_multitalk.py

@@ -6,6 +6,7 @@ import sys
 import json
 import warnings
 from datetime import datetime
+import gc
 
 warnings.filterwarnings('ignore')
 
@@ -494,6 +495,13 @@ def generate(args):
         
     logging.info("Finished.")
 
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()  # optional but useful with multiprocessing
+    gc.collect()
+
+    if dist.is_initialized():
+        dist.destroy_process_group()
 
 if __name__ == "__main__":
     args = _parse_args()