diff --git a/deepspeech/exps/u2/model.py b/deepspeech/exps/u2/model.py
index 8ab9a26e83590648695ade9563d498943360d602..2b6e2433068cac2d3cf28b2ca2ae87c9d7077af7 100644
--- a/deepspeech/exps/u2/model.py
+++ b/deepspeech/exps/u2/model.py
@@ -34,6 +34,7 @@ from deepspeech.io.sampler import SortagradDistributedBatchSampler
 from deepspeech.models.u2 import U2Model
 from deepspeech.training.optimizer import OptimizerFactory
 from deepspeech.training.scheduler import LRSchedulerFactory
+from deepspeech.training.timer import Timer
 from deepspeech.training.trainer import Trainer
 from deepspeech.utils import ctc_utils
 from deepspeech.utils import error_rate
@@ -184,40 +185,42 @@ class U2Trainer(Trainer):
             self.save(tag='init')
 
         self.lr_scheduler.step(self.iteration)
-        if self.parallel:
+        if self.parallel and hasattr(self.train_loader, 'batch_sampler'):
             self.train_loader.batch_sampler.set_epoch(self.epoch)
 
         logger.info(f"Train Total Examples: {len(self.train_loader.dataset)}")
         while self.epoch < self.config.training.n_epoch:
-            self.model.train()
-            try:
-                data_start_time = time.time()
-                for batch_index, batch in enumerate(self.train_loader):
-                    dataload_time = time.time() - data_start_time
-                    msg = "Train: Rank: {}, ".format(dist.get_rank())
-                    msg += "epoch: {}, ".format(self.epoch)
-                    msg += "step: {}, ".format(self.iteration)
-                    msg += "batch : {}/{}, ".format(batch_index + 1,
-                                                    len(self.train_loader))
-                    msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
-                    msg += "data time: {:>.3f}s, ".format(dataload_time)
-                    self.train_batch(batch_index, batch, msg)
+            with Timer("Epoch-Train Time Cost: {}"):
+                self.model.train()
+                try:
                     data_start_time = time.time()
-            except Exception as e:
-                logger.error(e)
-                raise e
-
-            total_loss, num_seen_utts = self.valid()
-            if dist.get_world_size() > 1:
-                num_seen_utts = paddle.to_tensor(num_seen_utts)
-                # the default operator in all_reduce function is sum.
-                dist.all_reduce(num_seen_utts)
-                total_loss = paddle.to_tensor(total_loss)
-                dist.all_reduce(total_loss)
-                cv_loss = total_loss / num_seen_utts
-                cv_loss = float(cv_loss)
-            else:
-                cv_loss = total_loss / num_seen_utts
+                    for batch_index, batch in enumerate(self.train_loader):
+                        dataload_time = time.time() - data_start_time
+                        msg = "Train: Rank: {}, ".format(dist.get_rank())
+                        msg += "epoch: {}, ".format(self.epoch)
+                        msg += "step: {}, ".format(self.iteration)
+                        msg += "batch : {}/{}, ".format(batch_index + 1,
+                                                        len(self.train_loader))
+                        msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
+                        msg += "data time: {:>.3f}s, ".format(dataload_time)
+                        self.train_batch(batch_index, batch, msg)
+                        data_start_time = time.time()
+                except Exception as e:
+                    logger.error(e)
+                    raise e
+
+            with Timer("Eval Time Cost: {}"):
+                total_loss, num_seen_utts = self.valid()
+                if dist.get_world_size() > 1:
+                    num_seen_utts = paddle.to_tensor(num_seen_utts)
+                    # the default operator in all_reduce function is sum.
+                    dist.all_reduce(num_seen_utts)
+                    total_loss = paddle.to_tensor(total_loss)
+                    dist.all_reduce(total_loss)
+                    cv_loss = total_loss / num_seen_utts
+                    cv_loss = float(cv_loss)
+                else:
+                    cv_loss = total_loss / num_seen_utts
 
             logger.info(
                 'Epoch {} Val info val_loss {}'.format(self.epoch, cv_loss))
diff --git a/deepspeech/exps/u2_kaldi/model.py b/deepspeech/exps/u2_kaldi/model.py
index 140ee947fae5eaa13e31f5002c3b6af5e3577c1a..095dfe34d67edbcdd54fd5ecb10a6a7305487342 100644
--- a/deepspeech/exps/u2_kaldi/model.py
+++ b/deepspeech/exps/u2_kaldi/model.py
@@ -32,6 +32,7 @@ from deepspeech.io.dataloader import BatchDataLoader
 from deepspeech.models.u2 import U2Model
 from deepspeech.training.optimizer import OptimizerFactory
 from deepspeech.training.scheduler import LRSchedulerFactory
+from deepspeech.training.timer import Timer
 from deepspeech.training.trainer import Trainer
 from deepspeech.utils import ctc_utils
 from deepspeech.utils import error_rate
@@ -190,35 +191,37 @@ class U2Trainer(Trainer):
 
         logger.info(f"Train Total Examples: {len(self.train_loader.dataset)}")
         while self.epoch < self.config.training.n_epoch:
-            self.model.train()
-            try:
-                data_start_time = time.time()
-                for batch_index, batch in enumerate(self.train_loader):
-                    dataload_time = time.time() - data_start_time
-                    msg = "Train: Rank: {}, ".format(dist.get_rank())
-                    msg += "epoch: {}, ".format(self.epoch)
-                    msg += "step: {}, ".format(self.iteration)
-                    msg += "batch : {}/{}, ".format(batch_index + 1,
-                                                    len(self.train_loader))
-                    msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
-                    msg += "data time: {:>.3f}s, ".format(dataload_time)
-                    self.train_batch(batch_index, batch, msg)
+            with Timer("Epoch-Train Time Cost: {}"):
+                self.model.train()
+                try:
                     data_start_time = time.time()
-            except Exception as e:
-                logger.error(e)
-                raise e
-
-            total_loss, num_seen_utts = self.valid()
-            if dist.get_world_size() > 1:
-                num_seen_utts = paddle.to_tensor(num_seen_utts)
-                # the default operator in all_reduce function is sum.
-                dist.all_reduce(num_seen_utts)
-                total_loss = paddle.to_tensor(total_loss)
-                dist.all_reduce(total_loss)
-                cv_loss = total_loss / num_seen_utts
-                cv_loss = float(cv_loss)
-            else:
-                cv_loss = total_loss / num_seen_utts
+                    for batch_index, batch in enumerate(self.train_loader):
+                        dataload_time = time.time() - data_start_time
+                        msg = "Train: Rank: {}, ".format(dist.get_rank())
+                        msg += "epoch: {}, ".format(self.epoch)
+                        msg += "step: {}, ".format(self.iteration)
+                        msg += "batch : {}/{}, ".format(batch_index + 1,
+                                                        len(self.train_loader))
+                        msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
+                        msg += "data time: {:>.3f}s, ".format(dataload_time)
+                        self.train_batch(batch_index, batch, msg)
+                        data_start_time = time.time()
+                except Exception as e:
+                    logger.error(e)
+                    raise e
+
+            with Timer("Eval Time Cost: {}"):
+                total_loss, num_seen_utts = self.valid()
+                if dist.get_world_size() > 1:
+                    num_seen_utts = paddle.to_tensor(num_seen_utts)
+                    # the default operator in all_reduce function is sum.
+                    dist.all_reduce(num_seen_utts)
+                    total_loss = paddle.to_tensor(total_loss)
+                    dist.all_reduce(total_loss)
+                    cv_loss = total_loss / num_seen_utts
+                    cv_loss = float(cv_loss)
+                else:
+                    cv_loss = total_loss / num_seen_utts
 
             logger.info(
                 'Epoch {} Val info val_loss {}'.format(self.epoch, cv_loss))
diff --git a/deepspeech/exps/u2_st/model.py b/deepspeech/exps/u2_st/model.py
index ef5938b7712f459a73ab7944bbb80c5ddf1fc81c..8dca16540e8871f5bac541ec2cec837ab5f95a05 100644
--- a/deepspeech/exps/u2_st/model.py
+++ b/deepspeech/exps/u2_st/model.py
@@ -38,6 +38,7 @@ from deepspeech.io.sampler import SortagradDistributedBatchSampler
 from deepspeech.models.u2_st import U2STModel
 from deepspeech.training.gradclip import ClipGradByGlobalNormWithLog
 from deepspeech.training.scheduler import WarmupLR
+from deepspeech.training.timer import Timer
 from deepspeech.training.trainer import Trainer
 from deepspeech.utils import bleu_score
 from deepspeech.utils import ctc_utils
@@ -207,35 +208,37 @@ class U2STTrainer(Trainer):
 
         logger.info(f"Train Total Examples: {len(self.train_loader.dataset)}")
         while self.epoch < self.config.training.n_epoch:
-            self.model.train()
-            try:
-                data_start_time = time.time()
-                for batch_index, batch in enumerate(self.train_loader):
-                    dataload_time = time.time() - data_start_time
-                    msg = "Train: Rank: {}, ".format(dist.get_rank())
-                    msg += "epoch: {}, ".format(self.epoch)
-                    msg += "step: {}, ".format(self.iteration)
-                    msg += "batch : {}/{}, ".format(batch_index + 1,
-                                                    len(self.train_loader))
-                    msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
-                    msg += "data time: {:>.3f}s, ".format(dataload_time)
-                    self.train_batch(batch_index, batch, msg)
+            with Timer("Epoch-Train Time Cost: {}"):
+                self.model.train()
+                try:
                     data_start_time = time.time()
-            except Exception as e:
-                logger.error(e)
-                raise e
-
-            total_loss, num_seen_utts = self.valid()
-            if dist.get_world_size() > 1:
-                num_seen_utts = paddle.to_tensor(num_seen_utts)
-                # the default operator in all_reduce function is sum.
-                dist.all_reduce(num_seen_utts)
-                total_loss = paddle.to_tensor(total_loss)
-                dist.all_reduce(total_loss)
-                cv_loss = total_loss / num_seen_utts
-                cv_loss = float(cv_loss)
-            else:
-                cv_loss = total_loss / num_seen_utts
+                    for batch_index, batch in enumerate(self.train_loader):
+                        dataload_time = time.time() - data_start_time
+                        msg = "Train: Rank: {}, ".format(dist.get_rank())
+                        msg += "epoch: {}, ".format(self.epoch)
+                        msg += "step: {}, ".format(self.iteration)
+                        msg += "batch : {}/{}, ".format(batch_index + 1,
+                                                        len(self.train_loader))
+                        msg += "lr: {:>.8f}, ".format(self.lr_scheduler())
+                        msg += "data time: {:>.3f}s, ".format(dataload_time)
+                        self.train_batch(batch_index, batch, msg)
+                        data_start_time = time.time()
+                except Exception as e:
+                    logger.error(e)
+                    raise e
+
+            with Timer("Eval Time Cost: {}"):
+                total_loss, num_seen_utts = self.valid()
+                if dist.get_world_size() > 1:
+                    num_seen_utts = paddle.to_tensor(num_seen_utts)
+                    # the default operator in all_reduce function is sum.
+                    dist.all_reduce(num_seen_utts)
+                    total_loss = paddle.to_tensor(total_loss)
+                    dist.all_reduce(total_loss)
+                    cv_loss = total_loss / num_seen_utts
+                    cv_loss = float(cv_loss)
+                else:
+                    cv_loss = total_loss / num_seen_utts
 
             logger.info(
                 'Epoch {} Val info val_loss {}'.format(self.epoch, cv_loss))
diff --git a/deepspeech/models/u2.py b/deepspeech/models/u2.py
index a01766da64e24b00835264022f092240cacfc8f0..fd8f154716459d109bdec3cb482fdc41ffc27ee8 100644
--- a/deepspeech/models/u2.py
+++ b/deepspeech/models/u2.py
@@ -115,7 +115,8 @@ class U2BaseModel(nn.Layer):
                  ctc_weight: float=0.5,
                  ignore_id: int=IGNORE_ID,
                  lsm_weight: float=0.0,
-                 length_normalized_loss: bool=False):
+                 length_normalized_loss: bool=False,
+                 **kwargs):
         assert 0.0 <= ctc_weight <= 1.0, ctc_weight
 
         super().__init__()
@@ -661,9 +662,7 @@ class U2BaseModel(nn.Layer):
             xs, offset, required_cache_size, subsampling_cache,
             elayers_output_cache, conformer_cnn_cache)
 
-    # @jit.to_static([
-    #         paddle.static.InputSpec(shape=[1, None, feat_dim],dtype='float32'),  # audio feat, [B,T,D]
-    #     ])
+    # @jit.to_static
     def ctc_activation(self, xs: paddle.Tensor) -> paddle.Tensor:
         """ Export interface for c++ call, apply linear transform and log
             softmax before ctc
@@ -830,6 +829,7 @@ class U2Model(U2BaseModel):
         Returns:
             int, nn.Layer, nn.Layer, nn.Layer: vocab size, encoder, decoder, ctc
         """
+        # cmvn
         if configs['cmvn_file'] is not None:
             mean, istd = load_cmvn(configs['cmvn_file'],
                                    configs['cmvn_file_type'])
@@ -839,11 +839,13 @@ class U2Model(U2BaseModel):
         else:
             global_cmvn = None
 
+        # input & output dim
         input_dim = configs['input_dim']
         vocab_size = configs['output_dim']
         assert input_dim != 0, input_dim
         assert vocab_size != 0, vocab_size
 
+        # encoder
         encoder_type = configs.get('encoder', 'transformer')
         logger.info(f"U2 Encoder type: {encoder_type}")
         if encoder_type == 'transformer':
@@ -855,17 +857,21 @@ class U2Model(U2BaseModel):
         else:
             raise ValueError(f"not support encoder type:{encoder_type}")
 
+        # decoder
         decoder = TransformerDecoder(vocab_size,
                                      encoder.output_size(),
                                      **configs['decoder_conf'])
+
+        # ctc decoder and ctc loss
+        model_conf = configs['model_conf']
         ctc = CTCDecoder(
             odim=vocab_size,
             enc_n_units=encoder.output_size(),
             blank_id=0,
-            dropout_rate=0.0,
+            dropout_rate=model_conf['ctc_dropoutrate'],
             reduction=True,  # sum
             batch_average=True,  # sum / batch_size
-            grad_norm_type='instance')
+            grad_norm_type=model_conf['ctc_grad_norm_type'])
 
         return vocab_size, encoder, decoder, ctc
 
diff --git a/deepspeech/models/u2_st.py b/deepspeech/models/u2_st.py
index 7dae3745bb97ee17633fd63d3c586a1f6987b86a..6737a549d978bb1f74a518140ec4e5917328ad1a 100644
--- a/deepspeech/models/u2_st.py
+++ b/deepspeech/models/u2_st.py
@@ -413,26 +413,26 @@ class U2STBaseModel(nn.Layer):
         best_hyps = best_hyps[:, 1:]
         return best_hyps
 
-    @jit.to_static
+    # @jit.to_static
     def subsampling_rate(self) -> int:
         """ Export interface for c++ call, return subsampling_rate of the
             model
         """
         return self.encoder.embed.subsampling_rate
 
-    @jit.to_static
+    # @jit.to_static
     def right_context(self) -> int:
         """ Export interface for c++ call, return right_context of the model
         """
         return self.encoder.embed.right_context
 
-    @jit.to_static
+    # @jit.to_static
     def sos_symbol(self) -> int:
         """ Export interface for c++ call, return sos symbol id of the model
         """
         return self.sos
 
-    @jit.to_static
+    # @jit.to_static
     def eos_symbol(self) -> int:
         """ Export interface for c++ call, return eos symbol id of the model
         """
@@ -468,7 +468,7 @@ class U2STBaseModel(nn.Layer):
             xs, offset, required_cache_size, subsampling_cache,
             elayers_output_cache, conformer_cnn_cache)
 
-    @jit.to_static
+    # @jit.to_static
     def ctc_activation(self, xs: paddle.Tensor) -> paddle.Tensor:
         """ Export interface for c++ call, apply linear transform and log
             softmax before ctc
@@ -643,14 +643,16 @@ class U2STModel(U2STBaseModel):
             decoder = TransformerDecoder(vocab_size,
                                          encoder.output_size(),
                                          **configs['decoder_conf'])
+            # ctc decoder and ctc loss
+            model_conf = configs['model_conf']
             ctc = CTCDecoder(
                 odim=vocab_size,
                 enc_n_units=encoder.output_size(),
                 blank_id=0,
-                dropout_rate=0.0,
+                dropout_rate=model_conf['ctc_dropout_rate'],
                 reduction=True,  # sum
                 batch_average=True,  # sum / batch_size
-                grad_norm_type='instance')
+                grad_norm_type=model_conf['ctc_grad_norm_type'])
 
             return vocab_size, encoder, (st_decoder, decoder, ctc)
         else:
diff --git a/deepspeech/modules/loss.py b/deepspeech/modules/loss.py
index 399e84e29cf42dad4aaec99766c73ccc1b2ee4fa..023a19232a0629bc3b820189fa6e32e2777d9fff 100644
--- a/deepspeech/modules/loss.py
+++ b/deepspeech/modules/loss.py
@@ -36,16 +36,16 @@ class CTCLoss(nn.Layer):
             f"CTCLoss Loss reduction: {reduction}, div-bs: {batch_average}")
 
         # instance for norm_by_times
-        # batchsize for norm_by_batchsize
+        # batch for norm_by_batchsize
         # frame for norm_by_total_logits_len
-        assert grad_norm_type in ('instance', 'batchsize', 'frame', None)
+        assert grad_norm_type in ('instance', 'batch', 'frame', None)
         self.norm_by_times = False
         self.norm_by_batchsize = False
         self.norm_by_total_logits_len = False
         logger.info(f"CTCLoss Grad Norm Type: {grad_norm_type}")
         if grad_norm_type == 'instance':
             self.norm_by_times = True
-        if grad_norm_type == 'batchsize':
+        if grad_norm_type == 'batch':
             self.norm_by_times = True
         if grad_norm_type == 'frame':
             self.norm_by_total_logits_len = True
diff --git a/deepspeech/training/gradclip.py b/deepspeech/training/gradclip.py
index 6393197a2016e730efd3b2eabafc5cf791239c23..87b36acaeccd9fecffba48a7b0c6c61a3ff782b2 100644
--- a/deepspeech/training/gradclip.py
+++ b/deepspeech/training/gradclip.py
@@ -47,9 +47,10 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
             sum_square = layers.reduce_sum(square)
             sum_square_list.append(sum_square)
 
-            # debug log
-            logger.debug(
-                f"Grad Before Clip: {p.name}: {float(sum_square.sqrt()) }")
+            # debug log, not dump all since slow down train process
+            if i < 10:
+                logger.debug(
+                    f"Grad Before Clip: {p.name}: {float(sum_square.sqrt()) }")
 
         # all parameters have been filterd out
         if len(sum_square_list) == 0:
@@ -75,9 +76,10 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
             new_grad = layers.elementwise_mul(x=g, y=clip_var)
             params_and_grads.append((p, new_grad))
 
-            # debug log
-            logger.debug(
-                f"Grad After Clip: {p.name}: {float(new_grad.square().sum().sqrt())}"
-            )
+            # debug log, not dump all since slow down train process
+            if i < 10:
+                logger.debug(
+                    f"Grad After Clip: {p.name}: {float(new_grad.square().sum().sqrt())}"
+                )
 
         return params_and_grads
diff --git a/doc/src/deepspeech_architecture.md b/doc/src/deepspeech_architecture.md
index c4c102baead3dea78d76b2364cf551d34d5dff64..ffa37affe903c8a760bde4a213c4406e33ef1801 100644
--- a/doc/src/deepspeech_architecture.md
+++ b/doc/src/deepspeech_architecture.md
@@ -183,5 +183,3 @@ bash run.sh --stage 0 --stop_stage 2 --model_type offline --conf_path conf/deeps
 cd examples/aishell/s0
 bash run.sh --stage 3 --stop_stage 5 --model_type offline --conf_path conf/deepspeech2.yaml
 ```
-
-
diff --git a/examples/aishell/s1/conf/chunk_conformer.yaml b/examples/aishell/s1/conf/chunk_conformer.yaml
index 3e606788ef86eee601233fab2039923b1ee8cb34..6f8ae135f6210757208dde85cad85e5ee776f381 100644
--- a/examples/aishell/s1/conf/chunk_conformer.yaml
+++ b/examples/aishell/s1/conf/chunk_conformer.yaml
@@ -76,6 +76,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/aishell/s1/conf/conformer.yaml b/examples/aishell/s1/conf/conformer.yaml
index 4b1430c58848a5cac7303518021a6256b52d525d..a4248459c261d442a3e23e18cf538927cda5236a 100644
--- a/examples/aishell/s1/conf/conformer.yaml
+++ b/examples/aishell/s1/conf/conformer.yaml
@@ -71,6 +71,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s1/conf/chunk_conformer.yaml b/examples/librispeech/s1/conf/chunk_conformer.yaml
index 0de1aefee5c4397c74ed2b382ceb4471baadd9c5..92db20f6688daea729cec7000e8ccbe6497ecb03 100644
--- a/examples/librispeech/s1/conf/chunk_conformer.yaml
+++ b/examples/librispeech/s1/conf/chunk_conformer.yaml
@@ -76,6 +76,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s1/conf/chunk_transformer.yaml b/examples/librispeech/s1/conf/chunk_transformer.yaml
index f782a0373680077eab678375b3677c5cfb3340f2..e0bc3135e0a75509da68561079127793a73b41a8 100644
--- a/examples/librispeech/s1/conf/chunk_transformer.yaml
+++ b/examples/librispeech/s1/conf/chunk_transformer.yaml
@@ -69,6 +69,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s1/conf/conformer.yaml b/examples/librispeech/s1/conf/conformer.yaml
index 6d825f05b4b1376f8ef9eda8551acc515947dfb2..78be249cb72e2b72f158cdb5f0187f64097edcf3 100644
--- a/examples/librispeech/s1/conf/conformer.yaml
+++ b/examples/librispeech/s1/conf/conformer.yaml
@@ -72,6 +72,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s1/conf/transformer.yaml b/examples/librispeech/s1/conf/transformer.yaml
index bc2ec606106808e6db332fe5bf0766a01e9007a9..4aa7b9158430da22552e2d75ab1383670efe67e1 100644
--- a/examples/librispeech/s1/conf/transformer.yaml
+++ b/examples/librispeech/s1/conf/transformer.yaml
@@ -33,7 +33,7 @@ collator:
   keep_transcription_text: False
   sortagrad: True 
   shuffle_method: batch_shuffle
-  num_workers: 2
+  num_workers: 0
 
 
 # network architecture
@@ -67,6 +67,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s1/local/train.sh b/examples/librispeech/s1/local/train.sh
index c57946a617e73c37c842c53aab7dde2a95d9d2c6..17a9e28df47234d20f3395e5579e07955a449c8d 100755
--- a/examples/librispeech/s1/local/train.sh
+++ b/examples/librispeech/s1/local/train.sh
@@ -20,7 +20,7 @@ echo "using ${device}..."
 mkdir -p exp
 
 seed=10086
-if [ ${seed} != 0]; then
+if [ ${seed} != 0 ]; then
     export FLAGS_cudnn_deterministic=True
 fi
 
diff --git a/examples/librispeech/s2/conf/chunk_conformer.yaml b/examples/librispeech/s2/conf/chunk_conformer.yaml
index 0de1aefee5c4397c74ed2b382ceb4471baadd9c5..92db20f6688daea729cec7000e8ccbe6497ecb03 100644
--- a/examples/librispeech/s2/conf/chunk_conformer.yaml
+++ b/examples/librispeech/s2/conf/chunk_conformer.yaml
@@ -76,6 +76,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s2/conf/chunk_transformer.yaml b/examples/librispeech/s2/conf/chunk_transformer.yaml
index f782a0373680077eab678375b3677c5cfb3340f2..e0bc3135e0a75509da68561079127793a73b41a8 100644
--- a/examples/librispeech/s2/conf/chunk_transformer.yaml
+++ b/examples/librispeech/s2/conf/chunk_transformer.yaml
@@ -69,6 +69,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s2/conf/conformer.yaml b/examples/librispeech/s2/conf/conformer.yaml
index 955b6108b8b4861dbc85db97e07f821d3333600f..9a72741350219b3455388d73d354ba9004890530 100644
--- a/examples/librispeech/s2/conf/conformer.yaml
+++ b/examples/librispeech/s2/conf/conformer.yaml
@@ -72,6 +72,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/librispeech/s2/conf/transformer.yaml b/examples/librispeech/s2/conf/transformer.yaml
index 4c60913e937bbd95bcd391e98a2b69a1242f63a6..edf5b81dc3a0c5061f76921e50b8095a5515c9d3 100644
--- a/examples/librispeech/s2/conf/transformer.yaml
+++ b/examples/librispeech/s2/conf/transformer.yaml
@@ -58,6 +58,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/ted_en_zh/t0/conf/transformer.yaml b/examples/ted_en_zh/t0/conf/transformer.yaml
index 755e044614f921037bf16b8e65458ed15c85b3e3..1aad86d22f02afc68d9eba0c0cb76406873fbfc4 100644
--- a/examples/ted_en_zh/t0/conf/transformer.yaml
+++ b/examples/ted_en_zh/t0/conf/transformer.yaml
@@ -68,6 +68,8 @@ model:
     model_conf:
         asr_weight: 0.0
         ctc_weight: 0.0
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/ted_en_zh/t0/conf/transformer_joint_noam.yaml b/examples/ted_en_zh/t0/conf/transformer_joint_noam.yaml
index bc1f8890d9d784053ce1ed84c9cbb8675f39782b..0144c40d4bf0f2b66bafefbda41acbc1708be170 100644
--- a/examples/ted_en_zh/t0/conf/transformer_joint_noam.yaml
+++ b/examples/ted_en_zh/t0/conf/transformer_joint_noam.yaml
@@ -68,6 +68,8 @@ model:
     model_conf:
         asr_weight: 0.5
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/timit/s1/conf/transformer.yaml b/examples/timit/s1/conf/transformer.yaml
index eb191d0b25f703c4e95aa79b750fcdaf8962be41..c3b519968822b888693b8cfb302a3fb7ad4156a6 100644
--- a/examples/timit/s1/conf/transformer.yaml
+++ b/examples/timit/s1/conf/transformer.yaml
@@ -66,6 +66,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/tiny/s1/conf/chunk_confermer.yaml b/examples/tiny/s1/conf/chunk_confermer.yaml
index 96da3d9fb2f2671cc20b7454727ea1ec77ee0630..be2e82f9e675e484f28ef0aead5306e36eefbc33 100644
--- a/examples/tiny/s1/conf/chunk_confermer.yaml
+++ b/examples/tiny/s1/conf/chunk_confermer.yaml
@@ -76,6 +76,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/tiny/s1/conf/chunk_transformer.yaml b/examples/tiny/s1/conf/chunk_transformer.yaml
index 1adb91c46ceded8d646c8a028c73e8a74b8e8fed..93439a85782f8cb08cc3040ae3a80625d9572117 100644
--- a/examples/tiny/s1/conf/chunk_transformer.yaml
+++ b/examples/tiny/s1/conf/chunk_transformer.yaml
@@ -69,6 +69,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/tiny/s1/conf/conformer.yaml b/examples/tiny/s1/conf/conformer.yaml
index b40e77e37993990b51b9817b9b3a68be14a65618..9bb67c44e54a968d2fb6e891592aaf0d0a2e3db4 100644
--- a/examples/tiny/s1/conf/conformer.yaml
+++ b/examples/tiny/s1/conf/conformer.yaml
@@ -72,6 +72,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
diff --git a/examples/tiny/s1/conf/transformer.yaml b/examples/tiny/s1/conf/transformer.yaml
index fd5adbdee359e8bc46185a3836d6d91ad0160795..5127e8c66821a866db27e383a7b8aa91dbab0a0b 100644
--- a/examples/tiny/s1/conf/transformer.yaml
+++ b/examples/tiny/s1/conf/transformer.yaml
@@ -66,6 +66,8 @@ model:
     # hybrid CTC/attention
     model_conf:
         ctc_weight: 0.3
+        ctc_dropoutrate: 0.0
+        ctc_grad_norm_type: instance
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false