diff --git a/test/test_distributions.py b/test/test_distributions.py
index 2dde18cf213..43433b95fd3 100644
--- a/test/test_distributions.py
+++ b/test/test_distributions.py
@@ -8,7 +8,7 @@
 import pytest
 import torch
 from _utils_internal import get_available_devices
-from torch import nn
+from torch import nn, autograd
 from torchrl.data.tensordict.tensordict import _TensorDict
 from torchrl.modules import (
     TanhNormal,
@@ -211,6 +211,18 @@ def test_tanhtrsf(dtype):
     assert (some_big_number.sign() == ones.sign()).all()
 
 
+@pytest.mark.parametrize("dtype", [torch.float, torch.double])
+def test_tanhtrsf_grad(dtype):
+    torch.manual_seed(0)
+    trsf = SafeTanhTransform()
+    x = torch.randn(100, requires_grad=True)
+    y1 = trsf(x)
+    y2 = x.tanh()
+    g1 = autograd.grad(y1.sum(), x, retain_graph=True)[0]
+    g2 = autograd.grad(y2.sum(), x, retain_graph=True)[0]
+    torch.testing.assert_close(g1, g2)
+
+
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()
     pytest.main([__file__, "--capture", "no", "--exitfirst"] + unknown)
diff --git a/torchrl/csrc/pybind.cpp b/torchrl/csrc/pybind.cpp
index 2e6fa604e6f..e3af8c3128d 100644
--- a/torchrl/csrc/pybind.cpp
+++ b/torchrl/csrc/pybind.cpp
@@ -11,6 +11,7 @@
 #include <memory>
 
 #include "segment_tree.h"
+#include "utils.h"
 
 namespace py = pybind11;
 
@@ -20,4 +21,6 @@ PYBIND11_MODULE(_torchrl, m) {
 
   torchrl::DefineMinSegmentTree<float>("Fp32", m);
   torchrl::DefineMinSegmentTree<double>("Fp64", m);
+
+  m.def("safetanh", &safetanh, "Safe Tanh");
 }
diff --git a/torchrl/csrc/utils.h b/torchrl/csrc/utils.h
new file mode 100644
index 00000000000..3f2944e2897
--- /dev/null
+++ b/torchrl/csrc/utils.h
@@ -0,0 +1,13 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+// This source code is licensed under the MIT license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <torch/extension.h>
+#include <torch/torch.h>
+
+#include <iostream>
+
+torch::Tensor safetanh(torch::Tensor input) {
+  return torch::clamp(torch::tanh(input), -0.999999, 0.999999);
+}
diff --git a/torchrl/modules/distributions/continuous.py b/torchrl/modules/distributions/continuous.py
index 4c88c7e967d..5ca1ec4c26a 100644
--- a/torchrl/modules/distributions/continuous.py
+++ b/torchrl/modules/distributions/continuous.py
@@ -11,6 +11,7 @@
 from torch import distributions as D, nn
 from torch.distributions import constraints
 
+from torchrl._torchrl import safetanh
 from torchrl.modules.distributions.truncated_normal import (
     TruncatedNormal as _TruncatedNormal,
 )
@@ -85,9 +86,7 @@ class SafeTanhTransform(D.TanhTransform):
     """
 
     def _call(self, x: torch.Tensor) -> torch.Tensor:
-        eps = torch.finfo(x.dtype).eps
-        y = super()._call(x)
-        y.data.clamp_(-1 + eps, 1 - eps)
+        y = safetanh(x)
         return y
 
     def _inverse(self, y: torch.Tensor) -> torch.Tensor: