fix off-by-one npast during some instances of fast forwarding

Author: LostRuins

gpttype_adapter.cpp

@@ -3468,7 +3468,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     if (debugmode==1 && !is_quiet)
     {
         std::string outstr = "";
-        printf("\n\n[Debug: Dump Raw Input Tokens]\n");
+        printf("\n\n[Debug: Dump %d Raw Input Tokens]\n",embd_inp.size());
         outstr += get_tok_vec_str(embd_inp);
         printf("%s\n", RemoveBell(outstr).c_str());
     }
@@ -3615,7 +3615,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     if (debugmode==1 && !is_quiet)
     {
         std::string outstr = "";
-        // printf("\n[Debug: Dump Forwarded Input Tokens, format: %d]\n", file_format);
+        // printf("\n[Debug: Dump Forwarded Input Tokens]\n");
         // outstr += get_tok_vec_str(embd_inp);
         outstr += "\n\n[Debug: n_past="+std::to_string(n_past)+" Context Size = " + std::to_string(current_context_tokens.size()) + "]\n";
         outstr += get_tok_vec_str(current_context_tokens);

koboldcpp.py

@@ -4367,6 +4367,9 @@ def auto_set_backend_gui(manual_select=False):
     def on_picked_model_file(filepath):
         if filepath and (filepath.lower().endswith('.kcpps') or filepath.lower().endswith('.kcppt')):
             #load it as a config file instead
+            if filepath.lower().endswith('.kcpps'):
+                global runmode_untouched
+                runmode_untouched = False
             with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
                 dict = json.load(f)
                 import_vars(dict)

model_adapter.cpp

@@ -464,9 +464,10 @@ void print_tok_vec(std::vector<float> &embd)
 
     //fast forward the past based on identical tokens, stop once a divergence is noted
     int embd_inp_len = embd_inp.size();
+    int cur_ctx_len = current_context_tokens.size();
     bool fastforwardok = true;
 
-    for (int i = 0; i < current_context_tokens.size(); ++i)
+    for (int i = 0; i < cur_ctx_len; ++i)
     {
         if (current_context_tokens[i] == embd_inp[i])
         {
@@ -500,6 +501,10 @@ void print_tok_vec(std::vector<float> &embd)
             {
                 break;
             }
+            if ((i + 2) >= cur_ctx_len)
+            {
+                break;
+            }
         }
     }