Update handler.py

handler.py

@@ -5,7 +5,7 @@ from diffusers import (
     EulerAncestralDiscreteScheduler,
 )
 from typing import Dict, List, Any
-
+from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
 import qrcode
 import os
 import base64
@@ -52,11 +52,13 @@ def load_models():
         "lllyasviel/control_v11f1e_sd15_tile",
         torch_dtype=torch.float16,
     )
+    controlnet_tile.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
 
     controlnet_brightness = ControlNetModel.from_pretrained(
         "ioclab/control_v1p_sd15_brightness",
         torch_dtype=torch.float16,
     )
+    controlnet_brightness.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
 
     pipe = StableDiffusionControlNetPipeline.from_pretrained(
         MODEL_ID,
@@ -67,10 +69,12 @@ def load_models():
         torch_dtype=torch.float16,
         cache_dir="cache",
         # local_files_only=True,
-    ).to("cuda")
+    )
+    
+    pipe.to("cuda")
 
     pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
-    pipe.enable_xformers_memory_efficient_attention()
+    pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
     return pipe