Update app.py

app.py

@@ -13,7 +13,7 @@ from diffusers.utils import load_image
 from PIL import Image
 import requests
 import transformers
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoTokenizer, T5EncoderModel, BitsAndBytesConfig
 from translatepy import Translator
 
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
@@ -53,10 +53,12 @@ transformer = SD3Transformer2DModel.from_pretrained(
     torch_dtype=torch.float16,
 )
 
+quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+
 text_encoder_3 = T5EncoderModel.from_pretrained(
     repo, 
     subfolder="text_encoder_3", 
-    torch_dtype=torch.float16,
+    quantization_config=quantization_config,
 )
 
 tokenizer_3 = AutoTokenizer.from_pretrained(
@@ -65,23 +67,44 @@ tokenizer_3 = AutoTokenizer.from_pretrained(
     torch_dtype=torch.float16,
 )
 
+torch.set_float32_matmul_precision("high")
+
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+
 # Ensure model and scheduler are initialized in GPU-enabled function
 if torch.cuda.is_available():
     pipe = StableDiffusion3Pipeline.from_pretrained(
         repo, 
-        vae=vae, 
-        transformer=transformer, 
         tokenizer_3=tokenizer_3, 
         text_encoder_3=text_encoder_3, 
         torch_dtype=torch.float16).to("cuda")
     pipe2 = StableDiffusion3Img2ImgPipeline.from_pretrained(
         repo, 
-        vae=vae, 
-        transformer=transformer, 
         tokenizer_3=tokenizer_3, 
         text_encoder_3=text_encoder_3, 
         torch_dtype=torch.float16).to("cuda")
 
+pipe.set_progress_bar_config(disable=True)
+
+pipe.transformer.to(memory_format=torch.channels_last)
+pipe.vae.to(memory_format=torch.channels_last)
+
+pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
+pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
+
+
+pipe2.set_progress_bar_config(disable=True)
+
+pipe2.transformer.to(memory_format=torch.channels_last)
+pipe2.vae.to(memory_format=torch.channels_last)
+
+pipe2.transformer = torch.compile(pipe2.transformer, mode="max-autotune", fullgraph=True)
+pipe2.vae.decode = torch.compile(pipe2.vae.decode, mode="max-autotune", fullgraph=True)
+
+
 pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe.scheduler.config)
 pipe2.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe2.scheduler.config)