Ray train bugfix (#2458)

* fix nccl pg destroy warning

* update

* ray bugfix

src/axolotl/train.py

@@ -509,6 +509,7 @@ def train(
     # Save the trained model and cleanup
     save_trained_model(cfg, trainer, model, safe_serialization)
     create_model_card(cfg, trainer)
-    cleanup_distributed()
+    if not cfg.use_ray:
+        cleanup_distributed()
 
     return model, tokenizer, trainer