fix(realtime): don't kill the listen loop on receive() timeout

simple_websocket's receive(timeout=1) returns None when no frame arrives in
time; it does NOT mean the connection closed. The recv_loop treated None as a
disconnect, so any client that pauses sending for >1s (i.e. any silence longer
than one second, or a mic that only transmits while speaking) permanently
stopped the session from hearing the user -- barge-in and the next turn both
went dead while generation kept streaming out.

Also default --device to MPS when CUDA is unavailable, so CPU-only Mac paths
stop falling back to cpu.

eval_omni.py

@@ -99,7 +99,7 @@ def main():
     parser.add_argument('--temperature', default=0.7, type=float, help="Thinker生成温度")
     parser.add_argument('--top_p', default=0.85, type=float, help="nucleus采样阈值")
     parser.add_argument('--output_dir', default='./output_audio/', type=str, help="输出音频保存目录")
-    parser.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu', type=str, help="运行设备")
+    parser.add_argument('--device', default='cuda' if torch.cuda.is_available() else ('mps' if torch.backends.mps.is_available() else 'cpu'), type=str, help="运行设备")
     parser.add_argument('--audio_dir', default='./dataset/eval_omni/', type=str, help="测试音频目录")
     parser.add_argument('--image_dir', default='./dataset/eval_omni/', type=str, help="测试图像目录")
     parser.add_argument('--open_thinking', default=0, type=int, help="是否开启思考模式（0=否，1=是）（思考模式下禁用audio输出）")

webui/web_demo.py

@@ -383,9 +383,12 @@ def realtime(ws):
         while alive[0]:
             try:
                 data = ws.receive(timeout=1)
-                if data is None: alive[0] = False; break
+            except Exception:
+                alive[0] = False; break          # 只有真断开/异常才退出
+            # flask_sock 超时返回 None，并不表示连接关闭：客户端只要暂停送帧 >1s
+            # （例如助手说话期间没人出声），旧写法会在这里误判断线并永久停止监听。
+            if data is not None:
                 q.put(data)
-            except: alive[0] = False; break
 
     threading.Thread(target=recv_loop, daemon=True).start()
     try:
@@ -493,7 +496,7 @@ def init_model(args):
 if __name__ == '__main__':
     p = argparse.ArgumentParser()
     p.add_argument('--load_from', default='./', help='模型权重搜索目录；目录下可放多个 HF 格式模型，WebUI 会自动扫描并允许切换。')
-    p.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu', help='推理设备；CUDA 可用时默认 cuda。显存不足或排查环境问题时可改为 cpu。')
+    p.add_argument('--device', default='cuda' if torch.cuda.is_available() else ('mps' if torch.backends.mps.is_available() else 'cpu'), help='推理设备；优先 CUDA，其次 Apple Silicon 的 MPS，否则 CPU。')
     p.add_argument('--port', default=7860, type=int, help='WebUI 服务端口；端口被占用或需要同时启动多个实例时调整。')
     p.add_argument('--audio_chunk_frames', default=4, type=int, help='流式播放每次解码的 Mimi frame 数；默认 4 约 320ms。WebUI 播放卡顿时可调大到 8/12，低延迟优先时保持 4。')
     p.add_argument('--audio_overlap', default=2, type=int, help='分块 Mimi 解码的重叠帧数；默认 2 用于缓解块边界断裂。一般不需要调整，边界杂音明显时可适当增大。')