nit

app.py

@@ -29,15 +29,16 @@ from qwenimage.models.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
 dtype = torch.bfloat16
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
+exp = Qwen_FA3_AoT_int8()
+# exp = Qwen_FA3_AoT_fp8()
+exp.load()
+
 @spaces.GPU(duration=1500)
-def load_pipe():
-    exp = Qwen_FA3_AoT_int8()
-    # exp = Qwen_FA3_AoT_fp8()
-    exp.load()
+def optim_pipe(exp):
     exp.optimize()
     return exp.pipe
 
-pipe = load_pipe()
+pipe = optim_pipe(exp)
 
 
 MAX_SEED = np.iinfo(np.int32).max