diff --git a/mlx/backend/cuda/logsumexp.cu b/mlx/backend/cuda/logsumexp.cu
index 9e61b9ff0..a85b51583 100644
--- a/mlx/backend/cuda/logsumexp.cu
+++ b/mlx/backend/cuda/logsumexp.cu
@@ -54,7 +54,7 @@ __global__ void logsumexp(const T* in, T* out, int axis_size) {
     // https://github.com/NVIDIA/online-softmax
     normalizer = normalizer * softmax_exp(prevmax - maxval);
     for (int i = 0; i < N_READS; i++) {
-      normalizer = normalizer + softmax_exp(vals[i] - maxval);
+      normalizer = normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
     }
   }
 
diff --git a/mlx/backend/cuda/softmax.cu b/mlx/backend/cuda/softmax.cu
index 362c88431..2ff4464b0 100644
--- a/mlx/backend/cuda/softmax.cu
+++ b/mlx/backend/cuda/softmax.cu
@@ -57,7 +57,8 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
     normalizer = normalizer * softmax_exp(prevmax - maxval);
 #pragma unroll
     for (int i = 0; i < N_READS; i++) {
-      normalizer = normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
+      normalizer =
+          normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
     }
   }