llama : fix kv_cache `n` init (close #1903)

1 files changed, 2 insertions, 0 deletions

diff --git a/llama.cpp b/llama.cpp
index a50846f7..a2916b3e 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -886,6 +886,7 @@ static bool kv_cache_init(
     const int64_t n_elements = n_embd*n_mem;
 
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+    cache.n = 0;
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size;
@@ -904,6 +905,7 @@ static bool kv_cache_init(
     ggml_set_name(cache.k, "cache_k");
     ggml_set_name(cache.v, "cache_v");
 
+    (void) n_gpu_layers;
 #ifdef GGML_USE_CUBLAS
     if (n_gpu_layers > n_layer + 1) {
         ggml_cuda_assign_buffers_no_scratch(cache.v);