diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 5272c6589..27b4f5ec8 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -6840,7 +6840,7 @@ void ggml_init_cublas() {
                 alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
                 alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
                 alloc_prop.location.id = id;
-                CU_CHECK(cuMemGetAllocationGranularity(&g_device_caps[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_MINIMUM));
+                CU_CHECK(cuMemGetAllocationGranularity(&g_device_caps[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
             }
 #endif // !defined(GGML_USE_HIPBLAS)
             g_device_caps[id].vmm = !!device_vmm;