.

meta_ui.py

@@ -72,6 +72,9 @@ def chat(
     rep_pen, pres_pen, stop_raw,
     log_state
 ):
+    import threading
+    from queue import Queue, Empty
+
     stop = [s.strip() for s in stop_raw.split(",") if s.strip()] or None
     samp = {
         "max_new_tokens": int(max_new),
@@ -79,12 +82,31 @@ def chat(
         "top_p": top_p,
         "top_k": int(top_k),
         "repetition_penalty": rep_pen,
-        "presence_penalty":  pres_pen,
+        "presence_penalty": pres_pen,
         **({"stop": stop} if stop else {})
     }
-    out = backend(user, samp)
-    # 返回回答，同时把 log_state 原样带回（不刷新由 Interval 处理）
-    return out, log_state
+
+    result_q = Queue()
+
+    # 后台线程执行 backend 推理
+    def worker():
+        out = backend(user, samp)
+        result_q.put(out)
+
+    thread = threading.Thread(target=worker)
+    thread.start()
+
+    # 先返回提示
+    yield "⏳ 正在生成中...", log_state
+
+    # 每 0.1 秒轮询结果队列（避免阻塞 UI）
+    while thread.is_alive() or not result_q.empty():
+        try:
+            result = result_q.get(timeout=0.1)
+            yield result, log_state
+        except Empty:
+            continue
+
 
 # ───────────────────── Gradio UI ─────────────────────
 with gr.Blocks(title="调试界面") as demo: