nits

common/arg.cpp

@@ -1308,18 +1308,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     } else if (ex == LLAMA_EXAMPLE_TOKENIZE) {
         params.parse_special = true; // parse special tokens by default, like the old tokenize tool
     } else if (ex == LLAMA_EXAMPLE_TTS) {
-        params.n_ctx      = 4096;
-        params.n_batch    = 4096;
-        params.embedding  = true; // the talker backbone's hidden states are read via llama_get_embeddings_ith
-        params.out_file   = "output.wav";
-        params.tts_lang    = "english";
-        // reference defaults (qwen_tts Qwen3TTSForConditionalGeneration.generate()):
-        // top_k=50, top_p=1.0, temperature=0.9, repetition_penalty=1.05 over the whole
-        // generated history -- without the penalty, runs degenerate into immediate
-        // no-speech (EOS) or endless non-terminating tails
-        params.sampling.top_k          = 50;
-        params.sampling.top_p          = 1.0f;
-        params.sampling.temp           = 0.9f;
+        params.out_file = "output.wav";
         params.sampling.penalty_repeat = 1.05f;
         params.sampling.penalty_last_n = -1;
     }

src/llama-arch.cpp

@@ -1025,6 +1025,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_MINIMAX_M3:
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_KIMI_LINEAR:
+        case LLM_ARCH_QWEN3TTS:
             return false;
         default:
             return true;

tools/mtmd/mtmd-helper-gen.cpp

@@ -97,7 +97,7 @@ public:
             return 1;
         }
 
-        const std::string lang   = inp->lang ? inp->lang : "english";
+        const std::string lang   = (inp->lang && inp->lang[0]) ? inp->lang : "english";
         const llama_token c_lang = find_special_token(vocab, ("<|codec_language_" + lang + "|>").c_str());
         if (c_lang == LLAMA_TOKEN_NULL) {
             LOG_ERR("mtmd_helper_gen_audio: unknown language '%s'\n", lang.c_str());

tools/tts/tts.cpp

@@ -16,6 +16,22 @@
  * For contributors: please keep this code simple and easy to understand. Do not add unnecessary complexity. The goal is to have a simple CLI for testing TTS support.
  */
 
+struct tts_timings {
+    int64_t t_start_us = ggml_time_us();
+    int64_t t_last_us  = t_start_us;
+
+    void report(int n_frames) {
+        const int64_t t_now_us = ggml_time_us();
+        if (t_now_us - t_last_us < 2000000) {
+            return;
+        }
+        t_last_us = t_now_us;
+        const double t_elapsed_s = (t_now_us - t_start_us) / 1e6;
+        const double fps = t_elapsed_s > 0 ? n_frames / t_elapsed_s : 0.0;
+        LOG_INF("frames generated: %d, speed: %.2f frames/s\n", n_frames, fps);
+    }
+};
+
 static void print_usage(int, char ** argv) {
     LOG("\nexample usage:\n");
     LOG("\n    %s -m backbone.gguf -mm mmproj.gguf -p \"text to speak\" -o output.wav", argv[0]);
@@ -45,6 +61,9 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    // always enable embd, so that we can pass hidden states to the audio generation helper
+    params.embedding = true;
+
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -88,10 +107,12 @@ int main(int argc, char ** argv) {
     inp.top_k       = params.sampling.top_k;
     inp.top_p       = params.sampling.top_p;
     inp.out_type    = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
+    const int64_t t_prompt_start_us = ggml_time_us();
     if (gen.set_input(&inp) != 0) {
         LOG_ERR("set_input failed\n");
         return 1;
     }
+    LOG_INF("prompt eval took %.2f seconds\n", (ggml_time_us() - t_prompt_start_us) / 1e6);
 
     // codec_0 (backbone) EOS token: ordinary LLM sampling concern, kept out of the
     // model-agnostic audio-generation helper
@@ -115,6 +136,9 @@ int main(int argc, char ** argv) {
     int n_frames = 0;
     llama_token sampled = sample_codec0();
     const float * h_state = llama_get_embeddings_ith(lctx, -1);
+
+    tts_timings timings;
+
     for (; n_frames < max_new && sampled != codec_eos_tok; n_frames++) {
         const float * h_next = nullptr;
         if (gen.step(sampled, h_state, &h_next) != 0) {
@@ -123,6 +147,7 @@ int main(int argc, char ** argv) {
         }
         h_state = h_next;
         sampled = sample_codec0();
+        timings.report(n_frames + 1);
     }
 
     int32_t      sample_rate = 0;