fix: address review comments on Qwen3-TTS Gradio demo

- Fix SERVER_PID bug: background vllm-omni directly (not piped through
  tee) so $! captures the actual server PID, preventing orphaned
  processes on cleanup
- Fix PCM format: handle raw PCM responses (no WAV header) by decoding
  as int16 samples at 24 kHz
- Auto-detect task type: query /v1/models at startup and infer the task
  type from the model name, pre-selecting the correct radio button

Signed-off-by: lishunyang <lishunyang12@163.com>

Author: lishunyang12

examples/online_serving/qwen3_tts/gradio_demo.py

@@ -38,6 +38,27 @@
 TASK_TYPES = ["CustomVoice", "VoiceDesign", "Base"]
 
 
+def detect_task_type(api_base: str) -> str | None:
+    """Auto-detect the task type from the model loaded on the server."""
+    try:
+        with httpx.Client(timeout=10.0) as client:
+            resp = client.get(
+                f"{api_base}/v1/models",
+                headers={"Authorization": "Bearer EMPTY"},
+            )
+        if resp.status_code == 200:
+            data = resp.json()
+            models = data.get("data", [])
+            if models:
+                model_id = models[0].get("id", "")
+                for task in TASK_TYPES:
+                    if task.lower() in model_id.lower():
+                        return task
+    except Exception:
+        pass
+    return None
+
+
 def fetch_voices(api_base: str) -> list[str]:
     """Fetch available voices from the server."""
     try:
@@ -148,6 +169,12 @@ def generate_speech(
 
     # Decode audio response
     try:
+        if response_format == "pcm":
+            # PCM: raw 16-bit signed LE samples, no container header.
+            # Qwen3-TTS outputs at 24 kHz by default.
+            pcm_sample_rate = 24000
+            audio_np = np.frombuffer(resp.content, dtype=np.int16).astype(np.float32) / 32767.0
+            return (pcm_sample_rate, audio_np)
         audio_np, sample_rate = sf.read(io.BytesIO(resp.content))
         if audio_np.ndim > 1:
             audio_np = audio_np[:, 0]
@@ -189,6 +216,7 @@ def on_task_type_change(task_type: str):
 
 def build_interface(api_base: str):
     """Build the Gradio interface."""
+    detected_task = detect_task_type(api_base)
     voices = fetch_voices(api_base)
 
     css = """
@@ -199,7 +227,10 @@ def build_interface(api_base: str):
 
     with gr.Blocks(css=css, title="Qwen3-TTS Demo") as demo:
         gr.Markdown("# Qwen3-TTS Online Serving Demo")
-        gr.Markdown(f"**Server:** `{api_base}`")
+        server_info = f"**Server:** `{api_base}`"
+        if detected_task:
+            server_info += f"  |  **Detected task:** `{detected_task}`"
+        gr.Markdown(server_info)
 
         with gr.Row():
             # Left column: inputs
@@ -213,7 +244,7 @@ def build_interface(api_base: str):
                 with gr.Row():
                     task_type = gr.Radio(
                         choices=TASK_TYPES,
-                        value="CustomVoice",
+                        value=detected_task or "CustomVoice",
                         label="Task Type",
                         scale=2,
                     )

examples/online_serving/qwen3_tts/run_gradio_demo.sh

@@ -106,6 +106,9 @@ echo "=========================================="
 cleanup() {
     echo ""
     echo "Shutting down..."
+    if [ -n "$LOG_TAIL_PID" ]; then
+        kill "$LOG_TAIL_PID" 2>/dev/null || true
+    fi
     if [ -n "$SERVER_PID" ]; then
         echo "Stopping vLLM server (PID: $SERVER_PID)..."
         kill "$SERVER_PID" 2>/dev/null || true
@@ -133,8 +136,11 @@ vllm-omni serve "$MODEL" \
     --gpu-memory-utilization 0.9 \
     --trust-remote-code \
     --enforce-eager \
-    --omni 2>&1 | tee "$LOG_FILE" &
+    --omni > "$LOG_FILE" 2>&1 &
 SERVER_PID=$!
+# Tail the log so the user can follow startup progress
+tail -f "$LOG_FILE" 2>/dev/null &
+LOG_TAIL_PID=$!
 
 # Wait for server startup
 echo ""