Fix SoftReLU fused operator numerical stability (apache#17849) (apach…

…e#19391)

* fix numerically unstable fused softrelu op

* implement test for softrelu numerical stability

Co-authored-by: RuRo <ruro.ruro@ya.ru>

src/operator/fusion/fused_op-inl.h

@@ -566,7 +566,10 @@ __device__ inline DType sigmoid(const DType val) {
 
 template <typename DType>
 __device__ inline DType softrelu(const DType val) {
-  return logf(1 + expf(val));
+  // Avoid overflow of exp for large inputs.
+  // The threshold 20 is chosen such that softrelu(a) = a
+  // for a > 20 using floating precision.
+  return val > 20 ? val : logf(1 + expf(val));
 }
 
 template <typename DType>

tests/python/gpu/test_fusion.py

@@ -138,6 +138,9 @@ def announce_check(op_name):
     for act_type in ['relu', 'sigmoid', 'tanh', 'softrelu', 'softsign']:
         announce_check("Activation(act_type='{}')".format(act_type))
         check_fused_symbol(mx.sym.Activation(a, act_type=act_type), a=arr)
+        if act_type == 'softrelu':
+            # Check that softrelu implementation doesn't overflow on large inputs
+            check_fused_symbol(mx.sym.Activation(a, act_type=act_type), a=1000 * arr)
 
     # Cast requires dtype
     for dtype in ['float16', 'float32', 'float64', 'int32']: