diff --git a/docs/source/nn.rst b/docs/source/nn.rst
index 1281e0d359e..96c8a195e97 100644
--- a/docs/source/nn.rst
+++ b/docs/source/nn.rst
@@ -17,6 +17,7 @@ Operators for neural networks
         BatchNorm3d,
         COCOReader,
         CTCLoss,
+        RNNTLoss,
         CoinFlip,
         ConstantPad1d,
         ConstantPad2d,
diff --git a/oneflow/core/functional/impl/nn_functor.cpp b/oneflow/core/functional/impl/nn_functor.cpp
index 86cfc08fcf1..0b36ea1c0dc 100644
--- a/oneflow/core/functional/impl/nn_functor.cpp
+++ b/oneflow/core/functional/impl/nn_functor.cpp
@@ -1105,13 +1105,17 @@ class RNNTlossFunctor {
     CHECK_EQ_OR_RETURN(labels->dtype()->data_type() , DataType::kInt32)<<"labels must be int32";
     CHECK_EQ_OR_RETURN(act_lens->dtype()->data_type() , DataType::kInt32)<<"act_lens must be int32";
     CHECK_EQ_OR_RETURN(label_lens->dtype()->data_type(), DataType::kInt32)<<"label_lens must be int32";
+
+    CHECK_EQ_OR_RETURN(acts->shape()->NumAxes(),4)<<"the dim of acts must be 4";
+    CHECK_EQ_OR_RETURN(labels->shape()->NumAxes(),2)<<"the dim of labels must be 2";
+    CHECK_EQ_OR_RETURN(act_lens->shape()->NumAxes(),1)<<"the dim of act_lens must be 1";
+    CHECK_EQ_OR_RETURN(label_lens->shape()->NumAxes(),1)<<"the dim of label_lens must be 1";
+ 
     MutableAttrMap attrs;
 
     JUST(attrs.SetAttr<int32_t>("blank_label", blank_label));
     JUST(attrs.SetAttr<int32_t>("num_threads", num_threads));
 
-    // std::shared_ptr<one::Tensor> acts_trans = JUST(LogSoftmax(acts,-1));
-
     return OpInterpUtil::Dispatch<Tensor>(*op_, {acts, labels, act_lens, label_lens}, attrs);
   }
 
diff --git a/oneflow/user/ops/rnnt_op.cpp b/oneflow/user/ops/rnnt_op.cpp
index 20915e3d70b..ac48ae6006b 100644
--- a/oneflow/user/ops/rnnt_op.cpp
+++ b/oneflow/user/ops/rnnt_op.cpp
@@ -50,19 +50,6 @@ REGISTER_USER_OP("RNNTloss")
       return Maybe<void>::Ok();
     });
 
-// REGISTER_USER_OP("RNNTloss_grad")
-//     .Input("grads")
-//     .Input("dy")
-//     .Output("dx")
-//     .SetDataTypeInferFn([](user_op::InferContext* ctx) -> Maybe<void> {
-//       CHECK_EQ_OR_RETURN(ctx->InputDType("prob", 0), ctx->InputDType("dy", 0));
-//       *ctx->OutputDType("dx", 0) = ctx->InputDType("grads", 0);
-//       return Maybe<void>::Ok();
-//     })
-//     .SetGetSbpFn([](user_op::SbpContext* ctx) -> Maybe<void> {
-//       ctx->NewBuilder().PartialSum(ctx->inputs()).PartialSum(ctx->outputs()).Build();
-//       return Maybe<void>::Ok();
-//     });
 
 REGISTER_USER_OP_GRAD("RNNTloss")
     .SetGenBackwardOpConfFn([](const user_op::UserOpWrapper& op,
diff --git a/python/oneflow/nn/modules/loss.py b/python/oneflow/nn/modules/loss.py
index c48296404d5..6293b13b70d 100644
--- a/python/oneflow/nn/modules/loss.py
+++ b/python/oneflow/nn/modules/loss.py
@@ -690,12 +690,56 @@ def forward(
 
 
 class RNNTLoss(_Loss):
+    """The RNN Tranducer loss.
+    The documentation is referenced from:
+    https://github.com/HawkAaron/warp-transducer/tree/master/pytorch_binding
+
+    This loss introduces probabilistic sequence transduction system, based entirely
+    on RNNs, that is in principle able to transform any input sequence into any finite, 
+    discrete output sequence
+
+    Args:
+        blank (int, optional): blank label. Default :math:`0`.
+        reduction (string, optional): Specifies the reduction to apply to the output:
+            ``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
+            ``'mean'``: the output losses will be divided by the bathsize and. Default: ``'mean'``
+
+    Shape:
+        - acts: Tensor of size :math:`(B, T, U+1, V)`
+          B is the minibatch index 
+          T is the time index
+          U is the label sequence length (+1 means blank label prepanded)
+          V indexes over activations for each symbol in the alphabet
+        - labels: 2 dimensional Tensor containing all the targets of the batch with zero padded.
+        - act_lens: A 1-D Tensor of ints, the length of each label for each example in the minibatch.
+        - label_lens: A 1-D Tensor of ints, the length of each label for each example in the minibatch.
+
+    Reference:
+        A. Graves et al.: Sequence Transduction with Recurrent Neural Networks
+        http://www.cs.toronto.edu/~graves/icml_2012.pdf
+
+    For example:
+
+    .. code-block:: python
+
+        >>> import oneflow as flow
+        
+        >>> acts = flow.rand(2,2,3,5)
+        >>> labels = flow.tensor([[1, 2],[2,2]],dtype=flow.int)
+        >>> acts_length = flow.tensor([2,2],dtype=flow.int)
+        >>> label_length = flow.tensor([2,2],dtype=flow.int)
+        >>> rnnt = flow.nn.RNNTLoss(blank=0,reduction="mean")
+        >>> rnnt(acts,labels,acts_length,label_length)
+        tensor([5.1914], dtype=oneflow.float32)
+    """
+
     def __init__(
-        self, blank: int = 0, reduction: str = "mean", thread: int=0
+        self, blank: int = 0, reduction: str = "mean"
     ) -> None:
+        
         super(RNNTLoss, self).__init__(reduction)
         self.blank = blank
-        self.thread = thread
+        self.thread = 0
     
     def forward(
         self,
@@ -706,7 +750,6 @@ def forward(
     ) -> Tensor:
         if not acts.is_cuda:
             acts = flow._C.log_softmax(acts, -1)
-            print("cpu")
         loss = flow._C.RNNTloss(acts, labels, act_lens, label_lens, self.blank, self.thread)
         if self.reduction in ['sum', 'mean']:
             loss = loss.sum().unsqueeze(-1)