fix trainer.py:multistep_trainer args bug

Author: SolenoidWGT

ding/framework/middleware/functional/trainer.py

@@ -71,8 +71,8 @@ def _train(ctx: Union["OnlineRLContext", "OfflineRLContext"]):
 
         if ctx.train_data is None:  # no enough data from data fetcher
             return
-        data = ctx.train_data.to(policy._device)
-        train_output = policy.forward(data)
+        # data = ctx.train_data.to(policy._device)
+        train_output = policy.forward(ctx.train_data)
         nonlocal last_log_iter
         if ctx.train_iter - last_log_iter >= log_freq:
             loss = np.mean([o['total_loss'] for o in train_output])