diff --git a/ltp/fastltp.py b/ltp/fastltp.py
index 1f2a49da6..0c3265b55 100644
--- a/ltp/fastltp.py
+++ b/ltp/fastltp.py
@@ -26,6 +26,8 @@ def __init__(self, *args, **kwargs):
         import onnxruntime as rt
         so = rt.SessionOptions()
         so.graph_optimization_level = rt.GraphOptimizationLevel.ORT_ENABLE_EXTENDED
+
+        # fixme should auto detect
         providers = ['CPUExecutionProvider'] if self.device.type == 'cpu' else ['GPUExecutionProvider']
 
         onnx_path = os.path.join(self.cache_dir, "ltp.onnx")
@@ -100,9 +102,9 @@ def forward(self, *args, **kwargs):
 
     @no_gard
     def seg(self, inputs: List[str]):
-        length = [len(text) for text in inputs]
         tokenizerd = self.tokenizer.batch_encode_plus(inputs, padding=True)
         pretrained_inputs = {key: convert(value) for key, value in tokenizerd.items()}
+        length = np.sum(pretrained_inputs['attention_mask'], axis=-1) - 2
 
         # todo: io binding
         cls, hidden, seg = self.onnx.run(None, pretrained_inputs)
diff --git a/ltp/ltp.py b/ltp/ltp.py
index 1c810c1cc..c4a405c75 100644
--- a/ltp/ltp.py
+++ b/ltp/ltp.py
@@ -133,12 +133,17 @@ def sent_split(inputs: List[str], flag: str = "all", limit: int = 510):
 
     @no_gard
     def seg(self, inputs: List[str]):
-        length = torch.as_tensor([len(text) for text in inputs], device=self.device)
         tokenizerd = self.tokenizer.batch_encode_plus(inputs, return_tensors='pt', padding=True)
+
+        input_ids = tokenizerd['input_ids'].to(self.device)
+        attention_mask = tokenizerd['attention_mask'].to(self.device)
+        token_type_ids = tokenizerd['token_type_ids'].to(self.device)
+        length = torch.sum(attention_mask, dim=-1) - 2
+
         pretrained_output, *_ = self.model.pretrained(
-            input_ids=tokenizerd['input_ids'].to(self.device),
-            attention_mask=tokenizerd['attention_mask'].to(self.device),
-            token_type_ids=tokenizerd['token_type_ids'].to(self.device)
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids
         )
 
         # remove [CLS] [SEP]