add more perf stats

expose.cpp

@@ -274,6 +274,14 @@ extern "C"
     {
         return last_seed;
     }
+    int get_last_draft_success()
+    {
+        return last_draft_success;
+    }
+     int get_last_draft_failed()
+    {
+        return last_draft_failed;
+    }
     int get_total_gens() {
         return total_gens;
     }
@@ -281,6 +289,14 @@ extern "C"
     {
         return total_img_gens;
     }
+    int get_total_tts_gens()
+    {
+        return total_tts_gens;
+    }
+     int get_total_transcribe_gens()
+    {
+        return total_transcribe_gens;
+    }
     int get_last_stop_reason() {
         return (int)last_stop_reason;
     }

expose.h

@@ -245,4 +245,8 @@ extern int last_token_count;
 extern int last_seed;
 extern int total_gens;
 extern int total_img_gens;
+extern int total_tts_gens;
+extern int total_transcribe_gens;
+extern int last_draft_success;
+extern int last_draft_failed;
 extern stop_reason last_stop_reason;

gpttype_adapter.cpp

@@ -62,6 +62,8 @@ float last_eval_time = 0;
 int last_token_count = 0;
 int last_seed = -1;
 int total_gens = 0;
+int last_draft_success = 0;
+int last_draft_failed = 0;
 stop_reason last_stop_reason = stop_reason::INVALID;
 std::vector<std::string> generated_tokens;
 
@@ -4002,6 +4004,8 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     last_process_time = pt1;
     last_token_count = realnpredict;
     last_seed = kcpp_data->seed;
+    last_draft_failed = draft_failures;
+    last_draft_success = draft_successes;
     total_gens += 1;
     concat_output_mtx.lock();
     concat_output_reader_copy_res = concat_output;

koboldcpp.py

@@ -465,6 +465,11 @@ def init_library():
     handle.get_last_process_time.restype = ctypes.c_float
     handle.get_last_token_count.restype = ctypes.c_int
     handle.get_last_seed.restype = ctypes.c_int
+    handle.get_last_draft_success.restype = ctypes.c_int
+    handle.get_last_draft_failed.restype = ctypes.c_int
+    handle.get_total_img_gens.restype = ctypes.c_int
+    handle.get_total_tts_gens.restype = ctypes.c_int
+    handle.get_total_transcribe_gens.restype = ctypes.c_int
     handle.get_total_gens.restype = ctypes.c_int
     handle.get_last_stop_reason.restype = ctypes.c_int
     handle.abort_generate.restype = ctypes.c_bool
@@ -2422,12 +2427,16 @@ Enter Prompt:<br>
             lastc = handle.get_last_token_count()
             totalgens = handle.get_total_gens()
             totalimggens = handle.get_total_img_gens()
+            totalttsgens = handle.get_total_tts_gens()
+            totaltranscribegens = handle.get_total_transcribe_gens()
             stopreason = handle.get_last_stop_reason()
             lastseed = handle.get_last_seed()
+            lastdraftsuccess = handle.get_last_draft_success()
+            lastdraftfailed = handle.get_last_draft_failed()
             uptime = time.time() - start_time
             idletime = time.time() - last_req_time
             is_quiet = True if (args.quiet and args.debugmode != 1) else False
-            response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc, "last_seed":lastseed, "total_gens":totalgens, "stop_reason":stopreason, "total_img_gens":totalimggens, "queue":requestsinqueue, "idle":(0 if modelbusy.locked() else 1), "hordeexitcounter":exitcounter, "uptime":uptime, "idletime":idletime, "quiet":is_quiet}).encode())
+            response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc, "last_seed":lastseed, "last_draft_success":lastdraftsuccess, "last_draft_failed":lastdraftfailed, "total_gens":totalgens, "stop_reason":stopreason, "total_img_gens":totalimggens, "total_tts_gens":totalttsgens, "total_transcribe_gens":totaltranscribegens, "queue":requestsinqueue, "idle":(0 if modelbusy.locked() else 1), "hordeexitcounter":exitcounter, "uptime":uptime, "idletime":idletime, "quiet":is_quiet}).encode())
 
         elif self.path.endswith('/api/extra/generate/check'):
             if not self.secure_endpoint():
@@ -4860,20 +4869,6 @@ def unload_libs():
     if handle and dll_close:
         print("Unloading Libraries...")
         dll_close(handle._handle)
-        del handle.load_model
-        del handle.generate
-        del handle.new_token
-        del handle.get_stream_count
-        del handle.has_finished
-        del handle.get_last_eval_time
-        del handle.get_last_process_time
-        del handle.get_last_token_count
-        del handle.get_last_seed
-        del handle.get_total_gens
-        del handle.get_last_stop_reason
-        del handle.abort_generate
-        del handle.token_count
-        del handle.get_pending_output
         del handle
         handle = None
 

otherarch/tts_adapter.cpp

@@ -480,6 +480,8 @@ static int code_terminate_id = 151670;
 static int nthreads = 4;
 static int tts_max_len = 4096;
 
+int total_tts_gens = 0;
+
 bool ttstype_load_model(const tts_load_model_inputs inputs)
 {
     tts_is_quiet = inputs.quiet;
@@ -986,6 +988,7 @@ tts_generation_outputs ttstype_generate(const tts_generation_inputs inputs)
         last_generation_settings_audio_seed = inputs.audio_seed;
         last_generation_settings_speaker_seed = inputs.speaker_seed;
         last_generation_settings_prompt = std::string(inputs.prompt);
+        total_tts_gens += 1;
 
         return output;
     }

otherarch/whispercpp/whisper_adapter.cpp

@@ -28,6 +28,7 @@ static bool whisper_is_quiet = false;
 static whisper_context * whisper_ctx = nullptr;
 static std::string whisper_output_text = "";
 
+int total_transcribe_gens = 0;
 
 static bool is_wav_buffer(const std::string buf) {
     // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
@@ -279,5 +280,6 @@ whisper_generation_outputs whispertype_generate(const whisper_generation_inputs
     }
     output.text = whisper_output_text.c_str();
     output.status = 1;
+    total_transcribe_gens += 1;
     return output;
 }