diff --git a/mlx/backend/cuda/device/gemm_sm70.cuh b/mlx/backend/cuda/device/gemm_sm70.cuh
index f66b890750..63b38aee56 100644
--- a/mlx/backend/cuda/device/gemm_sm70.cuh
+++ b/mlx/backend/cuda/device/gemm_sm70.cuh
@@ -220,9 +220,9 @@ CUTE_DEVICE void gemm_sm70_mainloop(
 template <bool KMajor>
 inline constexpr auto make_matrix_stride(int m, int k) {
   if constexpr (KMajor) {
-    return cute::make_stride(k, cute::Int<1>{}, m * k);
+    return cute::make_stride(k, cute::Int<1>{}, int64_t(m) * k);
   } else {
-    return cute::make_stride(cute::Int<1>{}, m, m * k);
+    return cute::make_stride(cute::Int<1>{}, m, int64_t(m) * k);
   }
 }
 
diff --git a/mlx/backend/cuda/device/qmm_sm80.cuh b/mlx/backend/cuda/device/qmm_sm80.cuh
index 476c2339cf..5fea5b51b2 100644
--- a/mlx/backend/cuda/device/qmm_sm80.cuh
+++ b/mlx/backend/cuda/device/qmm_sm80.cuh
@@ -303,9 +303,9 @@ void qmm_sm80_kernel(
   int l_coord = int(blockIdx.z);
 
   // Define layouts (mixed).
-  auto dA = make_stride(k, Int<1>{}, m * k); // (dM,dK,dL)
-  auto dB = make_stride(k, Int<1>{}, n * k); // (dN,dK,dL)
-  auto dC = make_stride(n, Int<1>{}, m * n); // (dM,dN,dL)
+  auto dA = make_stride(k, Int<1>{}, int64_t(m) * k); // (dM,dK,dL)
+  auto dB = make_stride(k, Int<1>{}, int64_t(n) * k); // (dN,dK,dL)
+  auto dC = make_stride(n, Int<1>{}, int64_t(m) * n); // (dM,dN,dL)
   auto S_layout = make_scales_layout<GroupSize>(n, k, l);
 
   // Handle broadcasting.
@@ -388,9 +388,9 @@ void qmm_sm80_rhs_kernel(
 
   // Define layouts (mixed). All gathered rows form one matrix; each run of
   // equal rhs_indices inside the row tile reuses one expert tile.
-  auto dA = make_stride(k, Int<1>{}, m * k); // (dM,dK,dL)
-  auto dB = make_stride(k, Int<1>{}, n * k); // (dN,dK,dL)
-  auto dC = make_stride(n, Int<1>{}, m * n); // (dM,dN,dL)
+  auto dA = make_stride(k, Int<1>{}, int64_t(m) * k); // (dM,dK,dL)
+  auto dB = make_stride(k, Int<1>{}, int64_t(n) * k); // (dN,dK,dL)
+  auto dC = make_stride(n, Int<1>{}, int64_t(m) * n); // (dM,dN,dL)
   auto S_layout = make_scales_layout<GroupSize>(n, k, group_count);
 
   Tensor mA_mkl = make_tensor(make_gmem_ptr(A),        make_shape(m, k, 1), dA); // (M,K,L)
diff --git a/mlx/backend/cuda/quantized/qmm/qmm_sm90.cu b/mlx/backend/cuda/quantized/qmm/qmm_sm90.cu
index 7bc5d83437..546e1755db 100644
--- a/mlx/backend/cuda/quantized/qmm/qmm_sm90.cu
+++ b/mlx/backend/cuda/quantized/qmm/qmm_sm90.cu
@@ -124,10 +124,10 @@ void qmm_sm90(
   const char* tag = "[quantized_matmul]";
   auto [m, n, k, l, broadcast_b] = make_problem_shape(x, w, out);
 
-  auto dA = make_stride(int64_t(k), Int<1>{}, int64_t(m * k));
-  auto dB = make_stride(int64_t(k), Int<1>{}, int64_t(n * k));
-  auto dS = make_stride(Int<1>{}, int64_t(n), int64_t(n * k / group_size));
-  auto dD = make_stride(Int<1>{}, int64_t(n), int64_t(m * n));
+  auto dA = make_stride(int64_t(k), Int<1>{}, int64_t(m) * k);
+  auto dB = make_stride(int64_t(k), Int<1>{}, int64_t(n) * k);
+  auto dS = make_stride(Int<1>{}, int64_t(n), int64_t(n) * k / group_size);
+  auto dD = make_stride(Int<1>{}, int64_t(n), int64_t(m) * n);
   if (broadcast_b) {
     get<2>(dB) = 0;
     get<2>(dS) = 0;