mtmd: correct token order

src/llama-vocab.cpp

@@ -2347,6 +2347,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "_<EOT>"
                     || t.first == "<|end_of_text|>"
                     || t.first == "<end_of_utterance>" // smoldocling
+                    || t.first == "<｜end▁of▁sentence｜>" // deepseek-ocr
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

tools/mtmd/mtmd-cli.cpp

@@ -222,14 +222,18 @@ static std::string chat_add_and_format(mtmd_cli_context & ctx, common_chat_msg &
 
 static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
     bool add_bos = ctx.chat_history.empty();
-    auto formatted_chat = chat_add_and_format(ctx, msg);
-    LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
 
     mtmd_input_text text;
-    text.text          = formatted_chat.c_str();
+    text.text          = msg.content.c_str();
     text.add_special   = add_bos;
     text.parse_special = true;
 
+    if (!mtmd_is_deepseekocr(ctx.ctx_vision.get())) {
+        auto formatted_chat = chat_add_and_format(ctx, msg);
+        LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
+        text.text = formatted_chat.c_str();        
+    }
+
     if (g_is_interrupted) return 0;
 
     mtmd::input_chunks chunks(mtmd_input_chunks_init());
@@ -332,6 +336,11 @@ int main(int argc, char ** argv) {
         }
 
     } else {
+        if (mtmd_is_deepseekocr(ctx.ctx_vision.get())) {
+            LOG_ERR("\n DeepSeek-OCR doesn't support chat mode.");
+            return 1;
+        }
+        
         LOG("\n Running in chat mode, available commands:");
         if (mtmd_support_vision(ctx.ctx_vision.get())) {
             LOG("\n   /image <path>    load an image");

tools/mtmd/mtmd.cpp

@@ -864,6 +864,10 @@ int mtmd_get_audio_bitrate(mtmd_context * ctx) {
     return 16000; // 16kHz
 }
 
+bool mtmd_is_deepseekocr(mtmd_context * ctx) {
+    return ctx->ctx_v && clip_is_deepseekocr(ctx->ctx_v);
+}
+
 //
 // public API functions
 //

tools/mtmd/mtmd.h

@@ -117,6 +117,9 @@ MTMD_API bool mtmd_support_audio(mtmd_context * ctx);
 // return -1 if audio is not supported
 MTMD_API int mtmd_get_audio_bitrate(mtmd_context * ctx);
 
+// whether the current model is DeepSeek-OCR
+MTMD_API bool mtmd_is_deepseekocr(mtmd_context * ctx);
+
 // mtmd_bitmap
 //
 // if bitmap is image: