tts : fix n_ubatch + make WavTokenizer cache-less (ggml-org#13713)

ggml-ci

Author: ggerganov

src/llama-model.cpp

@@ -13189,6 +13189,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_NOMIC_BERT_MOE:
+        case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 res = nullptr;
             } break;

tools/tts/tts.cpp

@@ -579,6 +579,8 @@ int main(int argc, char ** argv) {
 
     params.model = params.vocoder.model;
     params.embedding = true;
+    params.ctx_shift = false; // silence warning
+    params.n_ubatch = params.n_batch;
 
     common_init_result llama_init_cts = common_init_from_params(params);
 
@@ -1020,8 +1022,8 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
     }
     GGML_ASSERT(batch.n_tokens == n_codes);
 
-    if (llama_decode(ctx_cts, batch) != 0) {
-        LOG_ERR("%s: llama_decode() failed\n", __func__);
+    if (llama_encode(ctx_cts, batch) != 0) {
+        LOG_ERR("%s: llama_encode() failed\n", __func__);
         return 1;
     }