From cc06f1171ba61690992fcb7e15dd0beff30822ac Mon Sep 17 00:00:00 2001
From: Howard Su <howard0su@gmail.com>
Date: Sat, 1 Jul 2023 22:37:26 +0800
Subject: [PATCH 1/2] Fix crash of test-tokenizer-0 under Debug build

---
 llama.cpp | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/llama.cpp b/llama.cpp
index 049f73e4449cd..2820802fb7e10 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -165,8 +165,8 @@ struct llama_layer {
 };
 
 struct llama_kv_cache {
-    struct ggml_tensor * k;
-    struct ggml_tensor * v;
+    struct ggml_tensor * k = NULL;
+    struct ggml_tensor * v = NULL;
 
     struct ggml_context * ctx = NULL;
 
@@ -180,8 +180,12 @@ 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
     }
 };

From a75ddde952c44f11cd2c03481ee4176729bd824d Mon Sep 17 00:00:00 2001
From: Howard Su <howard0su@gmail.com>
Date: Mon, 3 Jul 2023 20:12:40 +0800
Subject: [PATCH 2/2] Change per comment

---
 ggml-cuda.cu | 2 +-
 llama.cpp    | 4 ----
 2 files changed, 1 insertion(+), 5 deletions(-)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 4e0d3dbdea4d4..8c4ef84e672ff 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 2820802fb7e10..93439624fee9d 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
     }
 };