fixed tts for outetts

koboldcpp.py

@@ -69,7 +69,7 @@ dry_seq_break_max = 128
 extra_images_max = 4 # for kontext/qwen img
 
 # global vars
-KcppVersion = "1.108.1"
+KcppVersion = "1.108.2"
 showdebug = True
 kcpp_instance = None #global running instance
 global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False, "restart_override_config_target":""}

otherarch/tts_adapter.cpp

@@ -1205,7 +1205,7 @@ static tts_generation_outputs ttstype_generate_outetts(const tts_generation_inpu
     else
     {
         // spectral operations
-        const int n_embd = llama_model_n_embd(model_cts);
+        const int n_embd = llama_model_n_embd_out(model_cts);
         const float * embd = llama_get_embeddings(cts_ctx);
         std::vector<float> audio = embd_to_audio(embd, n_codes, n_embd, nthreads);
 

otherarch/ttscpp/src/ttscpp.cpp

@@ -1,5 +1,6 @@
 #include "ttscpp.h"
 #include <mutex>
+#include "llama-impl.h"
 
 // A list of all of the top level GGUF names under kokoro.duration_predictor that have quantization compatible tensors.
 static constexpr std::array<const char *, 5> DURATION_PREDICTOR_QUANTIZATION_COMPATIBLE_PARTS = {
@@ -337,12 +338,6 @@ static void zeros(std::ofstream & file, size_t n) {
     }
 }
 
-template <typename T>
-struct do_no_init {
-    T value;
-    do_no_init() { /* do nothing */ }
-};
-
 void quantize_gguf(const std::string & ifile, const std::string & ofile, struct quantization_params * params) {
     ggml_context * weight_ctx = NULL;
     struct gguf_init_params gguf_params = {
@@ -376,7 +371,7 @@ void quantize_gguf(const std::string & ifile, const std::string & ofile, struct
         }
     }
 
-    std::vector<do_no_init<uint8_t>> work;
+    std::vector<no_init<uint8_t>> work;
 
     std::ofstream fout;
     auto close_ofstream = [&]() {