diff --git a/llama.cpp b/llama.cpp
index f4d290d47..720786206 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -13515,9 +13515,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
     gguf_set_val_u32(ctx_out, "general.file_type", ftype);
     // Remove split metadata
-    gguf_remove_key(ctx_out, "split.no");
-    gguf_remove_key(ctx_out, "split.count");
-    gguf_remove_key(ctx_out, "split.tensors.count");
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_NO).c_str());
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str());
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str());
 
     if (params->kv_overrides) {
         const std::vector<llama_model_kv_override> & overrides = *(const std::vector<llama_model_kv_override> *)params->kv_overrides;