add server_schema for tts

tools/server/README.md

@@ -796,7 +796,7 @@ Returns raw audio bytes (`audio/wav` by default) rather than JSON. For more info
 
 `lang`: Language code for the utterance (model-dependent, e.g. `en`, `zh`). Optional.
 
-`speaker_ref_b64`: Base64-encoded reference audio to clone the speaker's voice. Optional. Alternatively, upload the reference audio as a `speaker_ref` file field via `multipart/form-data` (see example below) — if both are provided, the uploaded file takes precedence.
+`speaker_ref_b64`: Base64-encoded reference audio to clone the speaker's voice. Optional. Alternatively, upload the reference audio as a `speaker_ref` file field via `multipart/form-data` (see example below) - if both are provided, the uploaded file takes precedence.
 
 `top_k`, `top_p`: Sampling params for the acoustic code predictor. Optional, model-dependent defaults apply.
 

tools/server/server-context.cpp

@@ -5246,8 +5246,8 @@ void server_routes::init_routes() {
 
         json body = json::parse(req.body);
 
-        // multipart form fields arrive as strings, correct the types of the typed params
-        for (const char * k : { "top_k", "n_predict" }) {
+        // multipart form fields arrive as strings, correct the types of the numeric params
+        for (const char * k : { "top_k", "n_predict", "max_tokens" }) {
             if (body.contains(k) && body[k].is_string()) {
                 body[k] = std::stoi(body[k].get<std::string>());
             }
@@ -5257,12 +5257,6 @@ void server_routes::init_routes() {
                 body[k] = std::stof(body[k].get<std::string>());
             }
         }
-        if (body.contains("seed") && body["seed"].is_string()) {
-            body["seed"] = (int64_t) std::stoll(body["seed"].get<std::string>());
-        }
-        if (body.contains("stream") && body["stream"].is_string()) {
-            body["stream"] = body["stream"] == "true";
-        }
 
         std::string prompt = json_value(body, "input", json_value(body, "prompt", std::string()));
         if (prompt.empty()) {
@@ -5271,25 +5265,19 @@ void server_routes::init_routes() {
         }
 
         const std::string response_format = json_value(body, "response_format", std::string("wav"));
-        const bool stream = json_value(body, "stream", false);
 
         server_task task(SERVER_TASK_TYPE_TTS);
+        server_schema::eval_tts_schema(params, task, body);
+        const bool stream = task.params.stream;
         task.tts_inp.set_prompt(prompt);
         task.tts_inp.set_lang(json_value(body, "lang", std::string()));
-        task.tts_inp.data.top_k     = json_value(body, "top_k", 0);
-        task.tts_inp.data.top_p     = json_value(body, "top_p", 0.0f);
         task.tts_inp.data.stream    = stream;
         task.tts_inp.data.out_type  = response_format == "pcm"
             ? MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM
             : MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
-        task.params.stream    = stream;
-        task.params.n_predict = json_value(body, "n_predict", -1);
-        task.params.sampling  = params.sampling; // baseline defaults, then apply overrides below
-        task.params.sampling.penalty_repeat  = json_value(body, "repeat_penalty", 1.05f);
         // note: -1 is clamped to 0 by the sampler, it will disable the penalty
-        task.params.sampling.penalty_last_n  = task.params.n_predict > 0 ? task.params.n_predict : 512;
-        task.params.sampling.seed = json_value(body, "seed", params.sampling.seed);
-        task.tts_inp.data.seed    = task.params.sampling.seed; // same seed for the codec/vocoder RNG
+        task.params.sampling.penalty_last_n = task.params.n_predict > 0 ? task.params.n_predict : 512;
+        task.params.sampling.seed = task.tts_inp.data.seed; // same seed for the codec/vocoder RNG
         if (task.tts_inp.data.top_k > 0) {
             task.params.sampling.top_k = task.tts_inp.data.top_k;
         }

tools/server/server-schema.cpp

@@ -566,6 +566,58 @@ task_params eval_llama_cmpl_schema(
     return params;
 }
 
+//
+// TTS schema
+//
+
+std::vector<std::unique_ptr<field>> make_tts_schema(server_task & task) {
+    std::vector<std::unique_ptr<field>> fields;
+    auto add = [&](field * f) {
+        fields.emplace_back(f);
+    };
+
+    add((new field_num("top_k", task.tts_inp.data.top_k))
+        ->set_limits(0, INT32_MAX)
+        ->set_desc("Top-k for the acoustic code predictor, 0 to use the model default"));
+
+    add((new field_num("top_p", task.tts_inp.data.top_p))
+        ->set_limits(0.0f, 1.0f)
+        ->set_desc("Top-p for the acoustic code predictor, 0.0 to use the model default"));
+
+    add((new field_num("seed", task.tts_inp.data.seed))
+        ->set_desc("RNG seed for the backbone sampler and the codec/vocoder (-1 = random)"));
+
+    add((new field_num("n_predict", task.params.n_predict))
+        ->set_hard_limits(-1, INT32_MAX)
+        ->add_alias("max_tokens")
+        ->set_desc("Max number of audio frames to generate"));
+
+    add((new field_num("repeat_penalty", task.params.sampling.penalty_repeat))
+        ->set_desc("Repetition penalty applied to the backbone sampler over the whole generation"));
+
+    add((new field_bool("stream", task.params.stream))
+        ->set_desc("Stream the audio as it becomes available instead of waiting for the full generation"));
+
+    return fields;
+}
+
+void eval_tts_schema(const common_params & params_base, server_task & task, const json & data) {
+    // baseline defaults, individual requests can override them
+    // TODO @ngxson : change the default values based on model
+    task.params.sampling = params_base.sampling;
+    task.params.sampling.penalty_repeat = 1.05f;
+    task.params.n_predict = -1;
+    task.tts_inp.data.top_k = 0;
+    task.tts_inp.data.top_p = 0.0f;
+    task.tts_inp.data.seed  = params_base.sampling.seed;
+
+    field_eval_context ctx(task.params);
+
+    for (const auto & f : make_tts_schema(task)) {
+        f->eval(ctx, data);
+    }
+}
+
 //
 // eval() implementations
 //

tools/server/server-schema.h

@@ -101,4 +101,12 @@ task_params eval_llama_cmpl_schema(
                     const std::vector<llama_logit_bias> & logit_bias_eog,
                     const json & data);
 
+std::vector<std::unique_ptr<field>> make_tts_schema(server_task & task);
+
+// evaluates the /tts request params into task.params and task.tts_inp.data
+void eval_tts_schema(
+                    const common_params & params_base,
+                    server_task & task,
+                    const json & data);
+
 } // namespace server_schema