server: use random media marker (#21962)

* server: use random media marker

* nits

* remove legacy <__image__> token

* revert special char in random

tools/mtmd/mtmd.cpp

@@ -109,7 +109,7 @@ mtmd_context_params mtmd_context_params_default() {
         /* use_gpu           */ true,
         /* print_timings     */ true,
         /* n_threads         */ 4,
-        /* image_marker      */ MTMD_DEFAULT_IMAGE_MARKER,
+        /* image_marker      */ nullptr,
         /* media_marker      */ mtmd_default_marker(),
         /* flash_attn_type   */ LLAMA_FLASH_ATTN_TYPE_AUTO,
         /* warmup            */ true,
@@ -169,7 +169,7 @@ struct mtmd_context {
         media_marker (ctx_params.media_marker),
         n_embd_text  (llama_model_n_embd_inp(text_model))
     {
-        if (std::string(ctx_params.image_marker) != MTMD_DEFAULT_IMAGE_MARKER) {
+        if (ctx_params.image_marker != nullptr) {
             throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
         }
 
@@ -584,9 +584,6 @@ struct mtmd_tokenizer {
         parse_special = text->parse_special;
         input_text    = text->text;
         vocab         = llama_model_get_vocab(ctx->text_model);
-
-        // for compatibility, we convert image marker to media marker
-        string_replace_all(input_text, MTMD_DEFAULT_IMAGE_MARKER, ctx->media_marker);
     }
 
     int32_t tokenize(mtmd_input_chunks * output) {

tools/mtmd/mtmd.h

@@ -46,9 +46,6 @@
 #    define MTMD_API
 #endif
 
-// deprecated marker, use mtmd_default_marker() instead
-#define MTMD_DEFAULT_IMAGE_MARKER "<__image__>"
-
 #ifdef __cplusplus
 extern "C" {
 #endif

tools/server/server-common.cpp

@@ -84,6 +84,14 @@ std::string gen_tool_call_id() {
     return random_string();
 }
 
+static std::string media_marker = "";
+const char * get_media_marker() {
+    if (media_marker.empty()) {
+        media_marker = "<__media_" + random_string() + "__>";
+    }
+    return media_marker.c_str();
+}
+
 //
 // lora utils
 //
@@ -975,7 +983,7 @@ json oaicompat_chat_params_parse(
                 handle_media(out_files, image_url, opt.media_path);
 
                 p["type"] = "media_marker";
-                p["text"] = mtmd_default_marker();
+                p["text"] = get_media_marker();
                 p.erase("image_url");
 
             } else if (type == "input_audio") {
@@ -996,7 +1004,7 @@ json oaicompat_chat_params_parse(
                 // TODO: add audio_url support by reusing handle_media()
 
                 p["type"] = "media_marker";
-                p["text"] = mtmd_default_marker();
+                p["text"] = get_media_marker();
                 p.erase("input_audio");
 
             } else if (type != "text") {
@@ -1460,7 +1468,7 @@ json convert_transcriptions_to_chatcmpl(
     if (!language.empty()) {
         prompt += string_format(" (language: %s)", language.c_str());
     }
-    prompt += mtmd_default_marker();
+    prompt += get_media_marker();
 
     json chatcmpl_body = inp_body; // copy all fields
     chatcmpl_body["messages"] = json::array({

tools/server/server-common.h

@@ -92,6 +92,9 @@ std::string random_string();
 std::string gen_chatcmplid();
 std::string gen_tool_call_id();
 
+// get a random marker; note: each time the server restarts, the marker will be different
+const char * get_media_marker();
+
 //
 // lora utils
 //

tools/server/server-context.cpp

@@ -708,6 +708,7 @@ private:
             mparams.warmup           = params_base.warmup;
             mparams.image_min_tokens = params_base.image_min_tokens;
             mparams.image_max_tokens = params_base.image_max_tokens;
+            mparams.media_marker     = get_media_marker();
 
             mctx = mtmd_init_from_file(mmproj_path.c_str(), model, mparams);
             if (mctx == nullptr) {