diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 51792794673bb..2cda14e143c17 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -1921,12 +1921,17 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
     if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
+
+        // with a [0, 2, 1, 3] perm. and ne02==1 the matrix strides need to be determined from dim 3:
+        const int64_t sma = ne02 == 1 ? nb03/nb00 : nb02/nb00;
+        const int64_t smb = ne02 == 1 ? s13       : s12;
+
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
-                       src1_ptr, cu_data_type_b, s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
+                alpha, src0_ptr, cu_data_type_a, nb01/nb00, sma,     // strideA
+                       src1_ptr, cu_data_type_b, s11,       smb,     // strideB
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0, // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));