.

meta_ui.py

@@ -152,31 +152,64 @@ def chat(
     # yield "⏳ 正在生成中...", log_state
 
 
-    while True:
-        # ⚠️ 线程已结束且队列已空 → 直接 return 让生成器终止
-        if not thread.is_alive() and result_q.empty():
-            break
-            # return  # ← 新增这一行
+    if api_suffix == "/v1/chat/completions":
+        while True:
+            if not thread.is_alive() and result_q.empty():
+                break
+            try:
+                result = result_q.get(timeout=0.1)
+            except Empty:
+                continue
 
-        try:
-            result = result_q.get(timeout=0.1)
-        except Empty:
-            continue
+            if isinstance(result, str):
+                result = {"text": result}
+            elif not isinstance(result, dict) or "text" not in result:
+                result = {"text": str(result)}
+
+            history.append({"role": "assistant", "content": result["text"]})
+            yield result["text"], history  # ✅ 显示模型输出，同时更新 history
+        return
+    else:
+        while thread.is_alive():
+            try:
+                result = result_q.get(timeout=0.1)
+                break
+            except Empty:
+                continue
 
-        # 统一格式
         if isinstance(result, str):
             result = {"text": result}
         elif not isinstance(result, dict) or "text" not in result:
             result = {"text": str(result)}
 
-        # yield result, log_state       # 第一次真正把模型回复丢给前端
-        if api_suffix == "/v1/chat/completions":
-            history.append({"role": "assistant", "content": result["text"]})
-            yield result["text"], None
-        else:
-            yield result["text"], log_state
+        yield result["text"], log_state  # ✅ 其它接口只输出文本，不更新 history
+        return
     
-       # return                       # ← 把旧的 break 换成 return
+        # while True:
+        #     # ⚠️ 线程已结束且队列已空 → 直接 return 让生成器终止
+        #     if not thread.is_alive() and result_q.empty():
+        #         break
+        #         # return  # ← 新增这一行
+
+        #     try:
+        #         result = result_q.get(timeout=0.1)
+        #     except Empty:
+        #         continue
+
+        #     # 统一格式
+        #     if isinstance(result, str):
+        #         result = {"text": result}
+        #     elif not isinstance(result, dict) or "text" not in result:
+        #         result = {"text": str(result)}
+
+        #     # yield result, log_state       # 第一次真正把模型回复丢给前端
+        #     if api_suffix == "/v1/chat/completions":
+        #         history.append({"role": "assistant", "content": result["text"]})
+        #         yield result["text"], None
+        #     else:
+        #         yield result["text"], log_state
+
+        # return                       # ← 把旧的 break 换成 return
 
 # ────────────────── Gradio UI ──────────────────
 with gr.Blocks(title="调试界面") as demo: