Update app.py

app.py

@@ -18,7 +18,7 @@ import paramiko
 import datetime
 #import diffusers
 from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, AutoencoderKL, EulerAncestralDiscreteScheduler
-from diffusers.models.attention_processor import AttnProcessor2_0
+from diffusers.models.attention_processor import Attention, AttnProcessor2_0
 from transformers import CLIPTextModelWithProjection, CLIPTextModel, CLIPTokenizer
 
 torch.backends.cuda.matmul.allow_tf32 = False
@@ -85,8 +85,12 @@ device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
 text_encoder = CLIPTextModel.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='text_encoder', token=True)#.to(device=device, dtype=torch.bfloat16)
 text_encoder_2 = CLIPTextModelWithProjection.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='text_encoder_2',token=True)#.to(device=device, dtype=torch.bfloat16)
-
 def load_and_prepare_model():
+    proc=Attention(
+        upcast_attention=True,
+      #  upcast_softmax = True,
+        processor = AttnProcessor2_0
+    )
     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)
@@ -119,7 +123,7 @@ def load_and_prepare_model():
     #pipe.unet.to(memory_format=torch.channels_last)
     pipe.to(device=device, dtype=torch.bfloat16)
     pipe.vae = vaeXL.to(device) #.to('cpu') #.to(torch.bfloat16)
-    pipe.unet.set_attn_processor(AttnProcessor2_0())
+    pipe.unet.set_attn_processor(proc)
     pipe.vae.set_default_attn_processor()
     return pipe