diff --git a/candle-nn/src/layer_norm.rs b/candle-nn/src/layer_norm.rs
index 26800a7b..23d0c01b 100644
--- a/candle-nn/src/layer_norm.rs
+++ b/candle-nn/src/layer_norm.rs
@@ -28,7 +28,7 @@
 //! ```
 //!
 //! [`Layer Normalization`]: https://arxiv.org/abs/1607.06450
-use candle::{DType, Result, Tensor, D};
+use candle::{DType, Module, Result, Tensor, D};
 
 #[derive(Debug, Clone, Copy, PartialEq)]
 pub struct LayerNormConfig {
@@ -105,7 +105,7 @@ impl LayerNorm {
     }
 }
 
-impl crate::Module for LayerNorm {
+impl Module for LayerNorm {
     fn forward(&self, x: &Tensor) -> Result<Tensor> {
         let x_dtype = x.dtype();
         let internal_dtype = match x_dtype {
@@ -162,11 +162,20 @@ impl RmsNorm {
     pub fn into_inner(self) -> LayerNorm {
         self.0
     }
+
+    /// Faster variant of the forward kernel, this can only be used on contiguous tensors though.
+    pub fn forward_diff(&self, xs: &Tensor) -> Result<Tensor> {
+        self.0.forward(xs)
+    }
 }
 
-impl crate::Module for RmsNorm {
+impl Module for RmsNorm {
     fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        self.0.forward(xs)
+        if xs.is_contiguous() {
+            crate::ops::rms_norm(xs, &self.0.weight, self.0.eps as f32)
+        } else {
+            self.0.forward(xs)
+        }
     }
 }
 
diff --git a/candle-transformers/src/models/with_tracing.rs b/candle-transformers/src/models/with_tracing.rs
index 1c34bfa2..f4706c7e 100644
--- a/candle-transformers/src/models/with_tracing.rs
+++ b/candle-transformers/src/models/with_tracing.rs
@@ -180,6 +180,11 @@ impl RmsNorm {
         let inner = candle_nn::rms_norm(size, eps, vb)?;
         Ok(Self { inner, span })
     }
+
+    pub fn forward_diff(&self, x: &Tensor) -> Result<Tensor> {
+        let _enter = self.span.enter();
+        self.inner.forward_diff(x)
+    }
 }
 
 impl Module for RmsNorm {