[feat] Add final save at the end (#6274)

* add final save

* default 1 episode

Author: TongLi3701

applications/ColossalChat/coati/distributed/consumer.py

@@ -119,7 +119,7 @@ def loop(self) -> None:
                     assert len(self.buffer) == 0
                     if self.lr_scheduler is not None:
                         self.lr_scheduler.step()
-                    if (step + 1) % self.save_interval == 0:
+                    if (step + 1) % self.save_interval == 0 or (step + 1) == self.num_update_per_episode:
                         if self.rank == 0:
                             print(f"Start saving policy model at step {step + 1}.")
                         save_path = os.path.join(self.save_dir, f"modeling-step-{step + 1}")

applications/ColossalChat/rl_example.py

@@ -107,7 +107,7 @@
         num_producers=args.num_inferencer,
         num_proc_per_producer=1,
         num_consumer_procs=args.num_trainers,
-        num_episodes=10,
+        num_episodes=1,
         inference_batch_size=args.inference_batch_size,
         inference_microbatch_size=args.inference_microbatch_size,
         train_batch_size=args.train_batch_size,