Update app.py

app.py

@@ -9,6 +9,16 @@ import numpy as np
 import random
 
 import torch
+torch.backends.cuda.matmul.allow_tf32 = False
+torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
+torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
+torch.backends.cudnn.allow_tf32 = False
+torch.backends.cudnn.deterministic = False
+torch.backends.cudnn.benchmark = False
+#torch.backends.cuda.preferred_blas_library="cublas"
+#torch.backends.cuda.preferred_linalg_library="cusolver"
+torch.set_float32_matmul_precision("highest")
+
 from diffusers import StableDiffusion3Pipeline, SD3Transformer2DModel, AutoencoderKL
 from transformers import CLIPTextModelWithProjection, T5EncoderModel
 from transformers import CLIPTokenizer, T5TokenizerFast
@@ -21,22 +31,8 @@ from image_gen_aux import UpscaleWithModel
 from huggingface_hub import hf_hub_download
 import datetime
 import cyper
-
-#from models.transformer_sd3 import SD3Transformer2DModel
-#from pipeline_stable_diffusion_3_ipa import StableDiffusion3Pipeline
-
 from PIL import Image
 
-torch.backends.cuda.matmul.allow_tf32 = False
-torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
-torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
-torch.backends.cudnn.allow_tf32 = False
-torch.backends.cudnn.deterministic = False
-torch.backends.cudnn.benchmark = False
-#torch.backends.cuda.preferred_blas_library="cublas"
-#torch.backends.cuda.preferred_linalg_library="cusolver"
-torch.set_float32_matmul_precision("highest")
-
 hftoken = os.getenv("HF_AUTH_TOKEN") 
 
 code = r'''
@@ -65,7 +61,8 @@ def upload_to_ftp(filename):
 pyx = cyper.inline(code, fast_indexing=True, directives=dict(boundscheck=False, wraparound=False, language_level=3))
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    #vae=AutoencoderKL.from_pretrained("ford442/sdxl-vae-bf16", use_safetensors=True, subfolder='vae',token=True)
+
+#vae=AutoencoderKL.from_pretrained("ford442/sdxl-vae-bf16", use_safetensors=True, subfolder='vae',token=True)
 vaeX=AutoencoderKL.from_pretrained("ford442/stable-diffusion-3.5-large-fp32", safety_checker=None, use_safetensors=True, subfolder='vae', low_cpu_mem_usage=False, torch_dtype=torch.float32, token=True)
 pipe = StableDiffusion3Pipeline.from_pretrained(
     #"stabilityai  #  stable-diffusion-3.5-large",
@@ -92,67 +89,14 @@ text_encoder_3=T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-larg
 ll_transformer=SD3Transformer2DModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='transformer',token=True).to(torch.device("cuda:0"), dtype=torch.bfloat16)
 pipe.transformer=ll_transformer
 pipe.load_lora_weights("ford442/sdxl-vae-bf16", weight_name="LoRA/UltraReal.safetensors")
-
 pipe.to(device=device, dtype=torch.bfloat16)
-#pipe.to(device)
-#pipe.vae=vaeX.to('cpu')
+
 upscaler_2 = UpscaleWithModel.from_pretrained("Kim2091/ClearRealityV1").to(torch.device('cpu'))
 
 MAX_SEED = np.iinfo(np.int32).max
 
 MAX_IMAGE_SIZE = 4096
 
-@spaces.GPU(duration=40)
-def infer_30(
-    prompt,
-    negative_prompt_1,
-    negative_prompt_2,
-    negative_prompt_3,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
-    progress=gr.Progress(track_tqdm=True),
-):
-    pipe.vae=vaeX.to('cpu')
-    pipe.transformer=ll_transformer
-    pipe.text_encoder=text_encoder #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True).to(device=device, dtype=torch.bfloat16)
-    pipe.text_encoder_2=text_encoder_2 #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True).to(device=device, dtype=torch.bfloat16)
-    pipe.text_encoder_3=text_encoder_3 #T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True).to(device=device, dtype=torch.bfloat16)
-    seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator(device='cuda').manual_seed(seed)
-    print('-- generating image --')
-    sd_image = pipe(
-            prompt=prompt,
-            prompt_2=prompt,
-            prompt_3=prompt,
-            negative_prompt=negative_prompt_1,
-            negative_prompt_2=negative_prompt_2,
-            negative_prompt_3=negative_prompt_3,
-            guidance_scale=guidance_scale,
-            num_inference_steps=num_inference_steps,
-            width=width,
-            height=height,
-         #   cross_attention_kwargs={"scale": 0.75},
-            generator=generator,
-            max_sequence_length=512
-    ).images[0]
-    print('-- got image --')
-    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
-    sd35_path = f"sd35ll_{timestamp}.png"
-    sd_image.save(sd35_path,optimize=False,compress_level=0)
-    pyx.upload_to_ftp(sd35_path)
-    #  pipe.unet.to('cpu')
-    upscaler_2.to(torch.device('cuda'))
-    with torch.no_grad():
-        upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
-    print('-- got upscaled image --')
-    downscale2 = upscale2.resize((upscale2.width // 4, upscale2.height // 4),Image.LANCZOS)
-    upscale_path = f"sd35ll_upscale_{timestamp}.png"
-    downscale2.save(upscale_path,optimize=False,compress_level=0)
-    pyx.upload_to_ftp(upscale_path)
-    return sd_image, prompt
-
 @spaces.GPU(duration=70)
 def infer_60(
     prompt,
@@ -184,7 +128,6 @@ def infer_60(
             num_inference_steps=num_inference_steps,
             width=width,
             height=height,
-         #   cross_attention_kwargs={"scale": 0.75},
             generator=generator,
             max_sequence_length=512
     ).images[0]
@@ -193,7 +136,6 @@ def infer_60(
     sd35_path = f"sd35ll_{timestamp}.png"
     sd_image.save(sd35_path,optimize=False,compress_level=0)
     pyx.upload_to_ftp(sd35_path)
-    #  pipe.unet.to('cpu')
     upscaler_2.to(torch.device('cuda'))
     with torch.no_grad():
         upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
@@ -235,7 +177,6 @@ def infer_90(
             num_inference_steps=num_inference_steps,
             width=width,
             height=height,
-         #   cross_attention_kwargs={"scale": 0.75},
             generator=generator,
             max_sequence_length=512
     ).images[0]
@@ -244,7 +185,6 @@ def infer_90(
     sd35_path = f"sd35ll_{timestamp}.png"
     sd_image.save(sd35_path,optimize=False,compress_level=0)
     pyx.upload_to_ftp(sd35_path)
-    #  pipe.unet.to('cpu')
     upscaler_2.to(torch.device('cuda'))
     with torch.no_grad():
         upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
@@ -255,8 +195,8 @@ def infer_90(
     pyx.upload_to_ftp(upscale_path)
     return sd_image, prompt
 
-@spaces.GPU(duration=110)
-def infer_100(
+@spaces.GPU(duration=120)
+def infer_110(
     prompt,
     negative_prompt_1,
     negative_prompt_2,
@@ -294,7 +234,6 @@ def infer_100(
     sd35_path = f"sd35ll_{timestamp}.png"
     sd_image.save(sd35_path,optimize=False,compress_level=0)
     pyx.upload_to_ftp(sd35_path)
-    #  pipe.unet.to('cpu')
     upscaler_2.to(torch.device('cuda'))
     with torch.no_grad():
         upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
@@ -322,10 +261,9 @@ with gr.Blocks(theme=gr.themes.Origin(),css=css) as demo:
                 placeholder="Enter your prompt",
                 container=False,
             )
-            run_button_30 = gr.Button("Run 30", scale=0, variant="primary")
             run_button_60 = gr.Button("Run 60", scale=0, variant="primary")
             run_button_90 = gr.Button("Run 90", scale=0, variant="primary")
-            run_button_100 = gr.Button("Run 100", scale=0, variant="primary")
+            run_button_110 = gr.Button("Run 110", scale=0, variant="primary")
         result = gr.Image(label="Result", show_label=False)
         with gr.Accordion("Advanced Settings", open=True):
             negative_prompt_1 = gr.Text(
@@ -382,21 +320,6 @@ with gr.Blocks(theme=gr.themes.Origin(),css=css) as demo:
                     value=50,
                 )
         gr.on(
-        triggers=[run_button_30.click, prompt.submit],
-        fn=infer_30,
-        inputs=[
-            prompt,
-            negative_prompt_1,
-            negative_prompt_2,
-            negative_prompt_3,
-            width,
-            height,
-            guidance_scale,
-            num_inference_steps,
-        ],
-        outputs=[result, expanded_prompt_output],
-        )
-        gr.on(
         triggers=[run_button_60.click, prompt.submit],
         fn=infer_60,
         inputs=[
@@ -427,8 +350,8 @@ with gr.Blocks(theme=gr.themes.Origin(),css=css) as demo:
         outputs=[result, expanded_prompt_output],
         )
         gr.on(
-        triggers=[run_button_100.click, prompt.submit],
-        fn=infer_100,
+        triggers=[run_button_110.click, prompt.submit],
+        fn=infer_110,
         inputs=[
             prompt,
             negative_prompt_1,