fix(sft_trainer): Fix global_tokens and total_tokens metrics always showing 0.0 (#4854)

Author: khazic

verl/trainer/sft_trainer.py

@@ -313,9 +313,9 @@ def fit(self):
                 data = tu.get_tensordict(tensor_dict=data, non_tensor_dict=meta_info)
                 batch_seqlens = self._get_batch_seqlens(data=data)
                 # this is necessary. Otherwise, it is interpreted as NonTensorStack
-                batch_seqlens = NonTensorData(batch_seqlens)
+                batch_seqlens_ntd = NonTensorData(batch_seqlens)
 
-                tu.assign_non_tensor(data, update_lr_scheduler=True, global_token_num=batch_seqlens)
+                tu.assign_non_tensor(data, update_lr_scheduler=True, global_token_num=batch_seqlens_ntd)
 
                 # start profile in SPMD mode
                 if global_step == self.start_profile_step:

verl/trainer/sft_trainer_ray.py

@@ -291,11 +291,11 @@ def fit(self):
                 global_step += 1
                 # construct tensordict
                 data = tu.get_tensordict(tensor_dict=data, non_tensor_dict=meta_info)
-                batch_seqlens = self._get_batch_seqlens(data=data)
+                batch_seqlens = self._get_batch_seqlens(data=data).tolist()
                 # this is necessary. Otherwise, it is interpreted as NonTensorStack
-                batch_seqlens = NonTensorData(batch_seqlens.tolist())
+                batch_seqlens_ntd = NonTensorData(batch_seqlens)
 
-                tu.assign_non_tensor(data, update_lr_scheduler=True, global_token_num=batch_seqlens)
+                tu.assign_non_tensor(data, update_lr_scheduler=True, global_token_num=batch_seqlens_ntd)
 
                 # start profile in SPMD mode
                 if global_step == self.start_profile_step: