Update app.py

app.py

@@ -73,6 +73,7 @@ _nodes_ready = False
 tasks_status = {}
 _last_used_architecture = None
 _last_used_unet = None
+# _previous_arch_for_gc = None 
 
 # ==========================================================================
 # ⚙️ Comfy Config
@@ -968,7 +969,7 @@ def build_workflow(
 
 # 🌟 增加全局变量用于追踪 GPU 域内的架构切换
 
-@spaces.GPU(duration=30)  # 🌟 降低 duration 到 30
+@spaces.GPU(duration=40)  # 🌟 降低 duration 到 40
 def generate(
     prompt: str, unet_choice: str, custom_w: int, custom_h: int,
     steps: int, cfg: float, seed: int, sampler: str, scheduler: str,
@@ -977,10 +978,24 @@ def generate(
     task_id: str = None,
 ) -> str:
 
-    global _gpu_last_used_arch, _gpu_last_used_unet
+    # global _previous_arch_for_gc
     
     print(f"[Token探针] task={task_id} | token_present={bool(LocalContext.request.get(None) and LocalContext.request.get(None).headers.get('x-ip-token'))}", flush=True)
-    
+
+    # 获取当前架构
+    # current_arch = UNET_CHOICES.get(unet_choice, {}).get("engine", "unknown")
+
+    # 🚫 【已注释】原因：经过极限测试，跨架构排队时会导致 ZeroGPU 重新申请显存，拖慢加载。
+    # 交给 ComfyUI 原生的 RAM_PRESSURE 机制（lru=8, ram=16.0）平滑过渡效果更好。
+    # if _previous_arch_for_gc is not None and _previous_arch_for_gc != current_arch:
+    #     import gc
+    #     if torch.cuda.is_available():
+    #         torch.cuda.empty_cache()
+    #         gc.collect()
+    #         print(f"[显存清理] 架构切换 {_previous_arch_for_gc} -> {current_arch}，已清理碎片", flush=True)
+    #
+    # _previous_arch_for_gc = current_arch
+
     if task_id:
         if task_id not in tasks_status:
             tasks_status[task_id] = {}
@@ -1217,9 +1232,14 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+from fastapi import HTTPException
 
 @app.post("/api/generate")
 async def handle_generate(req: GenerateRequest, background_tasks: BackgroundTasks, request: Request):
+    # 🌟 兜底拦截：单边最大 1368，防绕过
+    if req.custom_w > 1536 or req.custom_h > 1536:
+        raise HTTPException(status_code=400, detail="单边最大支持 1536 像素，请调整宽高。")
+    
     if req.char_enable and req.char_lora:
         if not req.char_lora.endswith(".safetensors"):
             req.char_lora = f"{req.char_lora}.safetensors"