added prefix for llava, reverted system role in template as it degreaded gemma3. truncated debug logs

gpttype_adapter.cpp

@@ -2796,11 +2796,12 @@ int GetThreadsToUse(bool blasmode)
 }
 
 //this function prepares the clip embds for llava. it's only needed when images change
-static void PrepareLlavaEmbds(const int nctx, const std::vector<int> & llava_sep)
+static void PrepareLlavaEmbds(const int nctx, const std::vector<int> & llava_sep, const std::vector<int> & llava_intro)
 {
     if(clp_ctx!=nullptr && clp_img_data!=nullptr)
     {
         int sepsize = llava_sep.size();
+        int introsize = llava_intro.size();
         last_llava_mem.clear();
 
         for(int i=0;i<llava_images.size();++i)
@@ -2829,6 +2830,10 @@ static void PrepareLlavaEmbds(const int nctx, const std::vector<int> & llava_sep
                 if(llava_images[i].clp_image_tokens>0 && llava_images[i].clp_image_tokens < nctx)
                 {
                     int tokcnt = (i==0?(llava_images[i].clp_image_tokens):(llava_images[i].clp_image_tokens+sepsize));
+                    if(i==0)
+                    {
+                        tokcnt += introsize;
+                    }
                     for(int n=0;n<tokcnt;++n)
                     {
                         last_llava_mem.push_back(current_llava_identifier);
@@ -3144,6 +3149,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     std::vector<int> embd_inp;
     std::vector<int> embd_inp_mem; //for storing added memory
     std::vector<int> llava_sep; //to separate between different llava images
+    std::vector<int> llava_intro; //to separate between different llava images
     bool llava_embds_built = false;
 
     int32_t nctx = kcpp_data->n_ctx;
@@ -3151,6 +3157,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     TokenizeString(kcpp_data->prompt, embd_inp, file_format, add_bos_token);
     bool use_mrope = (file_format == FileFormat::GGUF_GENERIC && file_format_meta.model_architecture == GGUFArch::ARCH_QWEN2VL);
     TokenizeString("\n\n", llava_sep, file_format, false);
+    TokenizeString("\nImages:\n", llava_intro, file_format, false);
 
     if(llava_composite_image_signature=="")
     {
@@ -3158,7 +3165,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     }
     if(llava_images_changed)
     {
-        PrepareLlavaEmbds(nctx, llava_sep);
+        PrepareLlavaEmbds(nctx, llava_sep, llava_intro);
         llava_embds_built = true;
     }
 
@@ -3872,7 +3879,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
                 {
                     if(!llava_embds_built) //this should never happen! however, handle it anyway
                     {
-                        PrepareLlavaEmbds(nctx, llava_sep);
+                        PrepareLlavaEmbds(nctx, llava_sep, llava_intro);
                         llava_embds_built = true;
                         printf("\nSomehow vision embd was not prepared (maybe no fast forward), rebuilding it...\n");
                     }
@@ -3888,6 +3895,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
                         int llavatokenscounted = 0;
                         int llavatokensevaled = 0;
                         int sepsize = llava_sep.size();
+                        int introsize = llava_intro.size();
                         while(input_consumed < embd_inp.size() && (embd_inp[input_consumed]==LLAVA_TOKEN_IDENTIFIER_A || embd_inp[input_consumed]==LLAVA_TOKEN_IDENTIFIER_B))
                         {
                             if (!last_n_tokens.empty())
@@ -3902,7 +3910,23 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
                         for(int i=0;i<llava_images.size();++i)
                         {
                             //note: no handling for draft_ctx as we don't support vision for it
-                            if(i>0 && sepsize>0)
+                            if(introsize>0 && i==0)
+                            {
+                                //added at the start of everything
+                                kcpp_embd_batch batch = kcpp_embd_batch(llava_intro, n_past, use_mrope, false);
+                                auto evr = llama_decode(llama_ctx_v4, batch.batch);
+                                if(evr!=0)
+                                {
+                                    printf("\nError when appending llava intro: %d\n",evr);
+                                }
+                                else
+                                {
+                                    printf("\rProcessing LLaVa Intro (%d tokens)",introsize);
+                                }
+                                n_past += introsize;
+                                llavatokensevaled += introsize;
+                            }
+                            if(sepsize>0 && i>0)
                             {
                                 //add a separator between each image
                                 kcpp_embd_batch batch = kcpp_embd_batch(llava_sep, n_past, use_mrope, false);

kcpp_adapters/AutoGuess.json

@@ -38,8 +38,6 @@
     "search": ["System role not supported", "<start_of_turn>"],
     "name": "Google Gemma 2.",
     "adapter": {
-        "system_start": "<start_of_turn>user\n",
-        "system_end": "<end_of_turn>\n",
         "user_start": "<start_of_turn>user\n",
         "user_end": "<end_of_turn>\n",
         "assistant_start": "<start_of_turn>model\n",
@@ -49,8 +47,6 @@
     "search": ["<start_of_image>", "<start_of_turn>", "<end_of_turn>"],
     "name": "Google Gemma 3.",
     "adapter": {
-        "system_start": "<start_of_turn>user\n",
-        "system_end": "<end_of_turn>\n",
         "user_start": "<start_of_turn>user\n",
         "user_end": "<end_of_turn>\n",
         "assistant_start": "<start_of_turn>model\n",

koboldcpp.py

@@ -49,7 +49,7 @@ logit_bias_max = 512
 dry_seq_break_max = 128
 
 # global vars
-KcppVersion = "1.87.3"
+KcppVersion = "1.87.4"
 showdebug = True
 kcpp_instance = None #global running instance
 global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False}
@@ -720,6 +720,22 @@ def string_contains_or_overlaps_sequence_substring(inputstr, sequences):
             return True
     return False
 
+def truncate_long_json(data, max_length):
+    if isinstance(data, dict):
+        new_data = {}
+        for key, value in data.items():
+            if isinstance(value, str):
+                new_data[key] = value[:max_length] + "..." if len(value) > max_length else value
+            else:
+                new_data[key] = truncate_long_json(value, max_length)
+        return new_data
+    elif isinstance(data, list):
+        return [truncate_long_json(item, max_length) for item in data]
+    elif isinstance(data, str):
+        return data[:max_length] + "..." if len(data) > max_length else data
+    else:
+        return data
+
 def get_capabilities():
     global savedata_obj, has_multiplayer, KcppVersion, friendlymodelname, friendlysdmodelname, fullsdmodelpath, mmprojpath, password, fullwhispermodelpath, ttsmodelpath, embeddingsmodelpath
     has_llm = not (friendlymodelname=="inactive")
@@ -2745,11 +2761,11 @@ Enter Prompt:<br>
         body = None
         if contlenstr:
             content_length = int(contlenstr)
-            if content_length > (1024*1024*32): #32mb payload limit
+            if content_length > (1024*1024*48): #48mb payload limit
                 self.send_response(500)
                 self.end_headers(content_type='application/json')
                 self.wfile.write(json.dumps({"detail": {
-                "msg": "Payload is too big. Max payload size is 32MB.",
+                "msg": "Payload is too big. Max payload size is 48MB.",
                 "type": "bad_input",
                 }}).encode())
                 return
@@ -2765,11 +2781,11 @@ Enter Prompt:<br>
                     if line:
                         chunk_length = max(0,int(line, 16))
                         content_length += chunk_length
-                    if not line or chunklimit > 512 or content_length > (1024*1024*32): #32mb payload limit
+                    if not line or chunklimit > 512 or content_length > (1024*1024*48): #48mb payload limit
                         self.send_response(500)
                         self.end_headers(content_type='application/json')
                         self.wfile.write(json.dumps({"detail": {
-                        "msg": "Payload is too big. Max payload size is 32MB.",
+                        "msg": "Payload is too big. Max payload size is 48MB.",
                         "type": "bad_input",
                         }}).encode())
                         return
@@ -3178,17 +3194,11 @@ Enter Prompt:<br>
                         }}).encode())
                         return
 
-
+                trunc_len = 8000
-                tmpimgs = genparams.get("images", []) # reduce amount of text printed to terminal when dumping large images
+                if args.debugmode >= 1:
-                if tmpimgs and isinstance(tmpimgs, (list, tuple)) and len(tmpimgs)>0:
+                    trunc_len = 16000
-                    printablegenparams = copy.deepcopy(genparams)
+                printablegenparams = truncate_long_json(genparams,trunc_len)
-                    outarr = []
+                utfprint("\nInput: " + json.dumps(printablegenparams),1)
-                    for img in tmpimgs:
-                        outarr.append(str(img[:512])+"...")
-                    printablegenparams["images"] = outarr
-                    utfprint("\nInput: " + json.dumps(printablegenparams),1)
-                else:
-                    utfprint("\nInput: " + json.dumps(genparams),1)
 
                 if args.foreground:
                     bring_terminal_to_foreground()