sequence_mask fix: when the input length is an empty tensor, the kernel tries...

sequence_mask fix: when the input length is an empty tensor, the kernel tries to dereference illegal sentinel iterator (#49525)

paddle/fluid/operators/sequence_ops/sequence_mask_op.h

@@ -106,7 +106,11 @@ class SequenceMaskKernel : public framework::OpKernel<Tx> {
 
     auto *x_data = x->data<Tx>();
     auto x_numel = x->numel();
+
     if (maxlen < 0) {
+      if (x_numel == 0) {
+        maxlen = 0;
+      } else {
 #if defined(__NVCC__) || defined(__HIPCC__)
         VLOG(10)
             << "SequenceMaskOp on GPU may be slow when maxlen is not provided.";
@@ -118,6 +122,7 @@ class SequenceMaskKernel : public framework::OpKernel<Tx> {
 #else
         maxlen = static_cast<int>(*std::max_element(x_data, x_data + x_numel));
 #endif
+      }
       auto y_dim = phi::vectorize<int>(x->dims());
       y_dim.push_back(maxlen);
       y->Resize(phi::make_ddim(y_dim));

paddle/fluid/operators/sequence_ops/sequence_mask_op_npu.cc

@@ -48,11 +48,15 @@ class SequenceMaskNPUKernel : public framework::OpKernel<T> {
 
     if (maxlen < 0) {
       auto x_numel = x->numel();
+      if (x_numel == 0) {
+        maxlen = 0;
+      } else {
         std::vector<T> x_vec;
         framework::TensorToVector(*x, dev_ctx, &x_vec);
         auto x_data = x_vec.data();
         maxlen = static_cast<int>(*std::max_element(x_data, x_data + x_numel));
       }
+    }
     auto y_dim = phi::vectorize<int>(x->dims());
     y_dim.push_back(maxlen);
 

python/paddle/fluid/tests/unittests/sequence/test_sequence_mask.py

@@ -17,6 +17,7 @@ import unittest
 
 import numpy as np
 
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.framework import (
     Program,
@@ -171,5 +172,14 @@ class TestSequenceMaskOpError(unittest.TestCase):
             self.assertRaises(TypeError, test_Variable)
 
 
+class TestSequenceMaskWithEmptyTensor(unittest.TestCase):
+    def test_empty(self):
+        paddle.disable_static()
+        lengths = paddle.to_tensor(np.array([], dtype=np.int64))
+        mask = paddle.nn.functional.sequence_mask(lengths)
+        self.assertEqual(list(mask.shape), [0, 0])
+        paddle.enable_static()
+
+
 if __name__ == '__main__':
     unittest.main()