Speed up rewrite: background short-video generation; topic_post text-only Qwen (no image input)

ai_ext.py

@@ -829,6 +829,33 @@ async def _generate_short_video(post, post_id: str, voice_id: str = None, speed:
         print(f"[short video error] {e}")
         return ""
 
+import threading as _threading
+def _spawn_background_video(post):
+    """Generate the short video in a BACKGROUND thread so the rewrite/topic endpoint
+    can return immediately. When done, persist the video URL onto the wall post.
+    This is the main fix for 'rewrite tu cac nguon tin qua lau' — the user gets the
+    text post instantly; the video appears shortly after (or via the 'Tao Video' button)."""
+    pid = post.get("id")
+    if not pid:
+        return
+    def _run():
+        try:
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+            video_url = loop.run_until_complete(_generate_short_video(post, pid))
+            loop.close()
+            if video_url:
+                posts = _load_wall()
+                for i, p in enumerate(posts):
+                    if str(p.get("id")) == str(pid):
+                        posts[i]["video"] = video_url
+                        break
+                _save_wall(posts)
+                print(f"[bg-video] done {pid} -> {video_url}")
+        except Exception as e:
+            print(f"[bg-video] error {pid}: {e}")
+    _threading.Thread(target=_run, daemon=True).start()
+
 # ===== MAKE POST =====
 def make_post(title, text, image, source_url, kind, sources=None, images=None, voice=None, emotion=None):
     # Auto-pick voice + emotion from the article topic when not provided
@@ -913,14 +940,12 @@ async def api_rewrite_share(request: Request):
                      images[0] if images else data.get("image", ""),
                      url, "rewrite", images=images)
 
-    # Auto-generate short video
-    video_url = await _generate_short_video(post, post["id"])
-    if video_url:
-        post["video"] = video_url
-
+    # Save post and return IMMEDIATELY; generate the short video in the background
+    # (so rewrite is fast). Video appears on the wall when ready / via 'Tao Video' button.
     posts = _load_wall()
     posts.insert(0, post)
     _save_wall(posts)
+    _spawn_background_video(post)
     return JSONResponse({"post": post})
 
 
@@ -948,13 +973,10 @@ async def api_url_wall(request: Request):
                      images[0] if images else data.get("image", ""),
                      url, "url", images=images)
 
-    video_url = await _generate_short_video(post, post["id"])
-    if video_url:
-        post["video"] = video_url
-
     posts = _load_wall()
     posts.insert(0, post)
     _save_wall(posts)
+    _spawn_background_video(post)
     return JSONResponse({"post": post})
 
 
@@ -971,19 +993,20 @@ async def api_topic_post(request: Request):
 
     image = pollinations_image_url(topic)
     prompt = _build_topic_prompt(topic, ctx)
-    text = await qwen_generate(prompt, image_url=image, max_tokens=500)
+    # NOTE: do NOT pass the decorative pollinations image to the VL model — feeding an
+    # image makes Qwen2.5-VL much slower (it must download+process it) with no benefit for
+    # a text summary. We keep the image only for display on the post. This is a major
+    # speed-up for 'rewrite tong hop'. (Text-only inference is several times faster.)
+    text = await qwen_generate(prompt, max_tokens=500)
     if not text:
         return JSONResponse({"error": "Qwen2.5-VL chưa sẵn sàng: " + LAST_QWEN_ERROR}, status_code=503)
     text = _clean_ai_output(text)
     post = make_post(topic, text, image, "", "topic")
 
-    video_url = await _generate_short_video(post, post["id"])
-    if video_url:
-        post["video"] = video_url
-
     posts = _load_wall()
     posts.insert(0, post)
     _save_wall(posts)
+    _spawn_background_video(post)
     return JSONResponse({"post": post})