Update app.py

app.py

@@ -123,7 +123,7 @@ def scheduler_swap_callback(pipeline, step_index, timestep, callback_kwargs):
     return {"latents": callback_kwargs["latents"]}   
 
 def load_and_prepare_model():
-    #vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", safety_checker=None, use_safetensors=False).to(device=device, dtype=torch.bfloat16)
+    vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", safety_checker=None, use_safetensors=False).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
     #vaeRV = AutoencoderKL.from_pretrained("SG161222/RealVisXL_V5.0", subfolder='vae', safety_checker=None, use_safetensors=False).to(device).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
     #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear")
     sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler')
@@ -132,10 +132,11 @@ def load_and_prepare_model():
         #torch_dtype=torch.bfloat16,
         add_watermarker=False,
       #  low_cpu_mem_usage = False,
-        token=HF_TOKEN,
+        token = HF_TOKEN,
+        scheduler = sched,
     )
-    #pipe.vae = vaeRV #.to(torch.bfloat16)
-    pipe.scheduler = sched
+    pipe.vae = vaeXL #.to(torch.bfloat16)
+    #pipe.scheduler = sched,
     #pipe.vae.do_resize=False
     #pipe.vae.vae_scale_factor=8
    # pipe.to(device=device, dtype=torch.bfloat16)
@@ -222,6 +223,7 @@ def scheduler_swap_callback(pipeline, step_index, timestep, callback_kwargs):
         torch.set_float32_matmul_precision("highest")
         #callback_kwargs["latents"] = callback_kwargs["latents"].to(torch.bfloat16)
         #pipe.unet.to(torch.float64)
+        pipe.unet.set_default_attn_processor()            ##   custom    ##
      #   pipe.vae = vae_a
      #   pipe.unet = unet_a
         torch.backends.cudnn.deterministic = False