diff --git a/tools/mtmd/mtmd-ios.cpp b/tools/mtmd/mtmd-ios.cpp
index 6c75e6b27bd..bf8de2ac1cb 100644
--- a/tools/mtmd/mtmd-ios.cpp
+++ b/tools/mtmd/mtmd-ios.cpp
@@ -149,6 +149,7 @@ void mtmd_ios_free(mtmd_ios_context* ctx) {
 }
 
 int mtmd_ios_prefill_image(mtmd_ios_context* ctx, const std::string& image_path) {
+
     if (!ctx || image_path.empty()) {
         return -1;
     }
@@ -292,3 +293,18 @@ mtmd_ios_token mtmd_ios_loop(mtmd_ios_context* ctx) {
 const char* mtmd_ios_get_last_error(mtmd_ios_context* ctx) {
     return ctx ? ctx->last_error.c_str() : nullptr;
 }
+
+bool mtmd_ios_clean_kv_cache(mtmd_ios_context* ctx) {
+    if (!ctx) {
+        return false;
+    }
+ 
+    // 清理 kv-cache 并重置序列位置
+    ctx->n_past = 0;
+    llama_kv_self_seq_rm(ctx->lctx, 0, 0, -1);
+
+    // 清理batch状态
+    common_batch_clear(ctx->batch);
+    
+    return true;
+}
\ No newline at end of file
diff --git a/tools/mtmd/mtmd-ios.h b/tools/mtmd/mtmd-ios.h
index b71453b4b9e..92a1f0e47c6 100644
--- a/tools/mtmd/mtmd-ios.h
+++ b/tools/mtmd/mtmd-ios.h
@@ -8,14 +8,10 @@
 #include "mtmd-helper.h"
 #include "mtmd.h"
 
-#ifdef __cplusplus
-extern "C" {
-#endif
-
 // Context structure
 typedef struct mtmd_ios_context mtmd_ios_context;
 
-// Parameters structure
+// Parameters structure (C++ only)
 typedef struct mtmd_ios_params {
     std::string model_path;
     std::string mmproj_path;
@@ -28,6 +24,16 @@ typedef struct mtmd_ios_params {
     bool        warmup;
 } mtmd_ios_params;
 
+// Loop return value structure (C++ only)
+typedef struct {
+    char * token;
+    bool   is_end;
+} mtmd_ios_token;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 // Initialize, returns 0 on success, -1 on failure
 // Parameters:
 // params: parameters
@@ -54,12 +60,6 @@ int mtmd_ios_prefill_image(mtmd_ios_context * ctx, const std::string & image_pat
 // role: role
 int mtmd_ios_prefill_text(mtmd_ios_context * ctx, const std::string & text, const std::string & role);
 
-// Loop return value structure
-typedef struct {
-    char * token;
-    bool   is_end;
-} mtmd_ios_token;
-
 // Loop, returns 0 on success, -1 on failure
 // Parameters:
 // ctx: context
@@ -75,6 +75,11 @@ const char * mtmd_ios_get_last_error(mtmd_ios_context * ctx);
 // str: string
 void mtmd_ios_string_free(char * str);
 
+// Clean kv-cache
+// Parameters:
+// ctx: context
+bool mtmd_ios_clean_kv_cache(mtmd_ios_context * ctx);
+
 #ifdef __cplusplus
 }
 #endif