Correct GLM-TTS model_specs_v1 metadata (0xShug0#123)

* Update GLM-TTS model spec v1 metadata

* Use normalized GLM-TTS request options

Author: mirek190

docs/community_models/glm_tts.md

@@ -89,11 +89,15 @@ audiocpp_cli --task clon --family glm_tts \
 | `--top-k` | integer | `25` | Speech-token top-k limit. |
 | `--top-p` | float | `0.8` | Speech-token nucleus threshold. |
 | `--seed` | integer | `0` | Seed used by token sampling, Flow noise, and HiFT. |
-| `--request-option flow_steps=<n>` | integer | `10` | Flow Euler integration steps. |
-| `--request-option cfg_rate=<float>` | float | `0.7` | Flow classifier-free guidance rate. |
-| `--request-option flow_noise_file=<path>` | raw F32 path | none | Optional exact initial Flow noise for parity tests. |
-| `--request-option hift_source_random_file=<path>` | raw F32 path | none | Optional exact HiFT phase-uniform and Gaussian values for parity tests. |
-| `--request-option hift_prior_noise_values=<n>` | integer | `0` | Torch RNG offset before normal HiFT source generation. |
+| `--request-option num_inference_steps=<n>` | integer | `10` | Flow Euler integration steps. |
+| `--request-option flow_guidance_scale=<float>` | float | `0.7` | Flow classifier-free guidance rate. |
+| `--request-option flow_noise_path=<path>` | raw F32 path | none | Optional exact initial Flow noise for parity tests. |
+| `--request-option hift_source_random_path=<path>` | raw F32 path | none | Optional exact HiFT phase-uniform and Gaussian values for parity tests. |
+| `--request-option hift_prior_noise_count=<n>` | integer | `0` | Torch RNG offset before normal HiFT source generation. |
+
+The legacy GLM-TTS request keys `flow_steps`, `cfg_rate`, `flow_noise_file`,
+`hift_source_random_file`, and `hift_prior_noise_values` remain accepted for
+backward compatibility.
 | `--session-option glm_tts.weight_type=native|f32|f16|bf16|q8_0` | enum | `native` | Requested component weight storage type. |
 | `--session-option glm_tts.mem_saver=true|false` | bool | `false` | Release the reference-only Whisper-VQ and CAMPPlus runtimes after caching the voice, while keeping Llama, Flow, and HiFT warm. |
 | `--session-option glm_tts.aggressive_mem_saver=true|false` | bool | `false` | Also release Llama, Flow, and HiFT after each stage. This minimizes VRAM but reloads the generation path on every request. |

docs/reports/glm_tts_validation.md

@@ -286,8 +286,8 @@ continue through the previous pooling graph. The regression probe
 `campplus_shared_default_probe` was compiled once against unmodified main and
 once against this branch; both emitted the same 192 serialized values.
 
-The request options `flow_noise_file`, `hift_source_random_file`, and
-`hift_prior_noise_values` expose stochastic boundaries for targeted parity
+The request options `flow_noise_path`, `hift_source_random_path`, and
+`hift_prior_noise_count` expose stochastic boundaries for targeted parity
 tests without adding normal-request log spam.
 
 ## Normal sampled-path parity

model_specs_v1/glm_tts.json

@@ -1,7 +1,7 @@
 {
   "family": "glm_tts",
   "display_name": "GLM-TTS",
-  "description": "Bilingual Chinese-English LLM-based TTS system with zero-shot voice cloning, streaming inference, reinforcement-learning-enhanced emotion and prosody control, and phoneme-level pronunciation control.",
+  "description": "Community Chinese-English zero-shot speech synthesis and voice cloning with native Llama, Whisper-VQ, Flow/DiT, CAMPPlus, and HiFT execution.",
   "category": "tts",
   "status": "community",
   "tasks": [
@@ -140,15 +140,15 @@
         "type": "int",
         "description": "Llama weight metadata context in MiB; default 8192.",
         "required": false,
-        "min": 0,
+        "min": 1,
         "default": 8192
       },
       {
         "name": "constant_context_mb",
         "type": "int",
         "description": "Llama constant tensor context in MiB; default 256.",
         "required": false,
-        "min": 0,
+        "min": 1,
         "default": 256
       }
     ],
@@ -228,15 +228,16 @@
   "packages": [
     {
       "id": "glm_tts_q8_0",
-      "display_name": "GLM-TTS Q8_0 GGUF",
+      "display_name": "GLM-TTS mixed Q8_0/F16 GGUF",
       "default": true,
       "format": "gguf",
       "precision": "q8_0",
-      "target_directory": "GLM-TTS-GGUF",
+      "target_directory": "GLM-TTS-Q8",
       "files": [
         "Text to audio (TTS)/GLM-TTS_Q8.gguf"
       ],
       "strip_prefix": "Text to audio (TTS)",
+      "description": "Standalone GGUF with the autoregressive Llama tensors retained in F16 and the speech tokenizer, Flow, HiFT, and CAMPPlus tensors stored as Q8_0.",
       "download": {
         "kind": "huggingface_snapshot",
         "repo": "mirek190/audio.cpp"

src/community_models/glm_tts/loader.cpp

@@ -49,20 +49,20 @@ runtime::ModelCliInterface cli() {
          "float",
          "Speech-token nucleus threshold; official default 0.8."},
         {"seed", "n", "Speech-token, Flow-noise, and HiFT seed."},
-        {"flow_steps",
+        {"num_inference_steps",
          "n",
          "Flow Euler steps; official default 10."},
-        {"cfg_rate",
+        {"flow_guidance_scale",
          "float",
          "Flow classifier-free guidance rate; official default 0.7."},
-        {"flow_noise_file",
+        {"flow_noise_path",
          "path",
          "Optional raw float32 initial Flow noise for parity tests."},
-        {"hift_source_random_file",
+        {"hift_source_random_path",
          "path",
          "Optional raw float32 HiFT phase-uniform and Gaussian values "
          "for parity tests."},
-        {"hift_prior_noise_values",
+        {"hift_prior_noise_count",
          "n",
          "Torch RNG value offset used before HiFT source generation."}};
     out.session_options = {

src/community_models/glm_tts/session.cpp

@@ -451,15 +451,15 @@ runtime::TaskResult GlmTTSSession::run(
     flow_input.speaker_embedding = reference.speaker_embedding;
     flow_input.inference_steps = static_cast<int>(
         runtime::parse_i64_option(
-            request.options, {"flow_steps"})
+            request.options, {"num_inference_steps", "flow_steps"})
             .value_or(assets_->config.flow.inference_steps));
     flow_input.cfg_rate = runtime::parse_finite_float_option(
-        request.options, {"cfg_rate"})
+        request.options, {"flow_guidance_scale", "cfg_rate"})
         .value_or(assets_->config.flow.inference_cfg_rate);
     const size_t flow_noise_count = static_cast<size_t>(
         flow_frames * assets_->config.flow.mel_dim);
     flow_input.initial_noise = optional_f32_values(
-        request, {"flow_noise_file"}, "Flow noise");
+        request, {"flow_noise_path", "flow_noise_file"}, "Flow noise");
     if (flow_input.initial_noise.empty()) {
         flow_input.initial_noise =
             normal_noise(flow_noise_count, seed + 1);
@@ -480,15 +480,16 @@ runtime::TaskResult GlmTTSSession::run(
         mel.mel, mel.frames, assets_->config.flow.mel_dim);
     const auto hift_source_random = optional_f32_values(
         request,
-        {"hift_source_random_file"},
+        {"hift_source_random_path", "hift_source_random_file"},
         "HiFT source-random");
     const auto * hift_source_random_ptr =
         hift_source_random.empty()
             ? nullptr
             : &hift_source_random;
     const uint64_t hift_prior_noise_values =
         runtime::parse_u64_option(
-            request.options, {"hift_prior_noise_values"})
+            request.options,
+            {"hift_prior_noise_count", "hift_prior_noise_values"})
             .value_or(0);
     const auto vocoder_start = Clock::now();
     auto waveform = hift().synthesize(

tests/glm_tts/glm_tts_warm_bench.cpp

@@ -135,8 +135,8 @@ std::vector<RequestCase> load_requests(
                  "top_k",
                  "top_p",
                  "seed",
-                 "flow_steps",
-                 "cfg_rate"}) {
+                 "num_inference_steps",
+                 "flow_guidance_scale"}) {
             if (const auto * value = item.find(name);
                 value != nullptr && !value->is_null()) {
                 request.options[name] = scalar_option(*value);