Update app.py

app.py

@@ -85,13 +85,13 @@ device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
 text_encoder = CLIPTextModel.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='text_encoder', token=True)#.to(device=device, dtype=torch.bfloat16)
 text_encoder_2 = CLIPTextModelWithProjection.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='text_encoder_2',token=True)#.to(device=device, dtype=torch.bfloat16)
-tokenizer_1 = CLIPTokenizer.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='tokenizer', token=True)
-tokenizer_2 = CLIPTokenizer.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='tokenizer_2', token=True)
-scheduler = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler', token=True)
-vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", low_cpu_mem_usage=False, safety_checker=None, use_safetensors=False, torch_dtype=torch.float32, token=True) #.to(device).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
-unet = UNet2DConditionModel.from_pretrained("ford442/RealVisXL_V5.0_BF16", low_cpu_mem_usage=False, subfolder='unet', upcast_attention=True, attention_type='gated-text-image', token=True)
 
 def load_and_prepare_model():
+    tokenizer_1 = CLIPTokenizer.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='tokenizer', token=True)
+    tokenizer_2 = CLIPTokenizer.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='tokenizer_2', token=True)
+    scheduler = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='scheduler', token=True)
+    vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", low_cpu_mem_usage=False, safety_checker=None, use_safetensors=False, torch_dtype=torch.float32, token=True) #.to(device).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
+    unet = UNet2DConditionModel.from_pretrained("ford442/RealVisXL_V5.0_BF16", low_cpu_mem_usage=False, subfolder='unet', upcast_attention=True, attention_type='gated-text-image', token=True)
     pipe = StableDiffusionXLPipeline.from_pretrained(
         'ford442/RealVisXL_V5.0_BF16',
         #torch_dtype=torch.bfloat16,
@@ -99,16 +99,16 @@ def load_and_prepare_model():
         add_watermarker=False,
         text_encoder=None,
         text_encoder_2=None,
-        #tokenizer=None,
-        #tokenizer_2=None,
-        #scheduler=None,
+        tokenizer=None,
+        tokenizer_2=None,
+        scheduler=None,
         unet=unet,
         vae=None,
     )
     #pipe.scheduler=scheduler
     #pipe.tokenizer=tokenizer_1
     #pipe.tokenizer_2=tokenizer_2
-    pipe.unet=unet
+    #pipe.unet=unet
     #pipe.vae.do_resize=False
     #pipe.vae.vae_scale_factor=8
     #pipe.to(device)