diff --git a/mlx/backend/cuda/binary_two.cu b/mlx/backend/cuda/binary_two.cu
index 3cdf2a89f..0a68e5f1d 100644
--- a/mlx/backend/cuda/binary_two.cu
+++ b/mlx/backend/cuda/binary_two.cu
@@ -153,7 +153,10 @@ void binary_op_gpu_inplace(
                     out_a.data_size() > INT32_MAX,
                 [&](auto large) {
                   using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-                  auto [shape, strides] = collapse_contiguous_dims(a, b, out_a);
+                  Shape shape;
+                  std::vector<Strides> strides;
+                  std::tie(shape, strides) =
+                      collapse_contiguous_dims(a, b, out_a);
                   auto& a_strides = strides[0];
                   auto& b_strides = strides[1];
                   int ndim = shape.size();
diff --git a/mlx/backend/cuda/ternary.cu b/mlx/backend/cuda/ternary.cu
index 06c063192..1d6535100 100644
--- a/mlx/backend/cuda/ternary.cu
+++ b/mlx/backend/cuda/ternary.cu
@@ -102,7 +102,9 @@ void ternary_op_gpu_inplace(
                 c.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
             [&](auto large) {
               using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-              auto [shape, strides] = collapse_contiguous_dims(a, b, c, out);
+              Shape shape;
+              std::vector<Strides> strides;
+              std::tie(shape, strides) = collapse_contiguous_dims(a, b, c, out);
               auto& a_strides = strides[0];
               auto& b_strides = strides[1];
               auto& c_strides = strides[2];