diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 4e0d3dbde..8c4ef84e6 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -2809,7 +2809,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
 }
 
 void ggml_cuda_free_data(struct ggml_tensor * tensor) {
-    if (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) {
+    if (!tensor || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
         return;
     }
 
diff --git a/llama.cpp b/llama.cpp
index 2820802fb..93439624f 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -180,12 +180,8 @@ struct llama_kv_cache {
         }
 
 #ifdef GGML_USE_CUBLAS
-    if (k) {
         ggml_cuda_free_data(k);
-    }
-    if (v) {
         ggml_cuda_free_data(v);
-    }
 #endif // GGML_USE_CUBLAS
     }
 };