Update app.py

app.py

@@ -12,8 +12,7 @@ import gradio as gr
 import numpy as np
 from PIL import Image
 import torch
-from diffusers import AutoencoderKL, StableDiffusionXLPipeline
-from diffusers import EulerAncestralDiscreteScheduler
+from diffusers import AutoencoderKL, StableDiffusionXLPipeline, UNet2DConditionModel, EulerAncestralDiscreteScheduler
 from typing import Tuple
 import paramiko
 import gc
@@ -101,9 +100,25 @@ def apply_style(style_name: str, positive: str, negative: str = "") -> Tuple[str
         negative = ""
     return p.replace("{prompt}", positive), n + negative
 
+model_repo='John6666/uber-realistic-porn-merge-xl-urpmxl-v6final-sdxl'
+        #'John6666/uber-realistic-porn-merge-ponyxl-urpm-ponyxlhybridv1-sdxl'
+        #'John6666/uber-realistic-porn-merge-xl-urpmxl-v3-sdxl'
+        #'John6666/uber-realistic-porn-merge-xl-urpmxl-v6final-sdxl'
+        #'John6666/pornworks-real-porn-ponyv04-sdxl'
+        #'John6666/pornmaster-amateur-sdxlv1vae-sdxl'
+
+
+rv='ford442/RealVisXL_V5.0_BF16'
+    
+text_encoder = CLIPTextModel.from_pretrained(rv, low_cpu_mem_usage=False, subfolder='text_encoder', token=True)#.to(device=device, dtype=torch.bfloat16)
+text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(model_repo, low_cpu_mem_usage=False, subfolder='text_encoder_2',token=True)#.to(device=device, dtype=torch.bfloat16)
+tokenizer_1 = CLIPTokenizer.from_pretrained(rv, low_cpu_mem_usage=False, subfolder='tokenizer', token=True)
+tokenizer_2 = CLIPTokenizer.from_pretrained(model_repo, low_cpu_mem_usage=False, subfolder='tokenizer_2', token=True)
+scheduler = EulerAncestralDiscreteScheduler.from_pretrained(rv, 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(model_repo, low_cpu_mem_usage=False, subfolder='unet', upcast_attention=True, attention_type='gated-text-image', token=True)
+
 def load_and_prepare_model():
-    vaeX = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", safety_checker=None,use_safetensors=False,torch_dtype=torch.float32,token=HF_TOKEN).to(device)
-    sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16',token=HF_TOKEN, subfolder='scheduler',beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012, steps_offset=1,use_karras_sigmas=False)
     pipe = StableDiffusionXLPipeline.from_pretrained(
         #'John6666/uber-realistic-porn-merge-ponyxl-urpm-ponyxlhybridv1-sdxl',
         #'John6666/uber-realistic-porn-merge-xl-urpmxl-v3-sdxl',
@@ -112,10 +127,21 @@ def load_and_prepare_model():
         #'John6666/pornmaster-amateur-sdxlv1vae-sdxl',
         #torch_dtype=torch.bfloat16,
         add_watermarker=False,
-        token=HF_TOKEN,
-        #use_safetensors=True,
+        token=True,
+        add_watermarker=False,
+        text_encoder=None,
+        text_encoder_2=None,
+        tokenizer=None,
+        tokenizer_2=None,
+        scheduler=None,
+        unet=None,
+        vae=None,
     )
    
+    pipe.scheduler=scheduler
+    pipe.tokenizer=tokenizer_1
+    pipe.tokenizer_2=tokenizer_2
+    pipe.unet=unet
     '''
     scaling_factor (`float`, *optional*, defaults to 0.18215):
             The component-wise standard deviation of the trained latent space computed using the first batch of the
@@ -136,10 +162,8 @@ def load_and_prepare_model():
     #****  NEVER TO BFLOAT BEFORE CUDA****#
 
     #pipe.vae.force_upcast=True
-    pipe.scheduler = sched
     # pipe.load_lora_weights("ford442/sdxl-vae-bf16", weight_name="LoRA/Fantasy_World_XL.safetensors", adapter_name="fantasy")
     pipe.vae.set_default_attn_processor()  # Set attention processor first
-
     pipe.load_lora_weights("ford442/sdxl-vae-bf16", weight_name="LoRA/skin_texture_style_v4.safetensors", adapter_name="skin", low_cpu_mem_usage=False,token=HF_TOKEN)
     
     #pipe.unet = pipe.unet.to(memory_format=torch.contiguous_format)
@@ -153,7 +177,8 @@ def load_and_prepare_model():
     #**** BETTER WAY ****#
     pipe.to(device, torch.bfloat16)
     #**** BETTER WAY ****#
-    pipe.vae = vaeX #.to(torch.bfloat16)
+    pipe.vae = vaeX.to(device) #.to(torch.bfloat16)
+    pipe.vae.set_default_attn_processor()
     #pipe.to(device)
     #pipe.to(torch.bfloat16)
 
@@ -198,19 +223,13 @@ def randomize_seed_fn() -> int:
 def uploadNote(prompt,num_inference_steps,guidance_scale,timestamp):
     filename= f'tst_A_{timestamp}.txt'
     with open(filename, "w") as f:
-        f.write(f"Realvis 5.0 (Tester A) \n")
+        f.write(f"Realvis 5.0 (Tester E) \n")
         f.write(f"Date/time: {timestamp} \n")
         f.write(f"Prompt: {prompt} \n")
         f.write(f"Steps: {num_inference_steps} \n")
         f.write(f"Guidance Scale: {guidance_scale} \n")
         f.write(f"SPACE SETUP: \n")
-        f.write(f"Use Model Dtype: no \n")
-        f.write(f"Model Scheduler: Euler_a all_custom before cuda \n")
-        f.write(f"Model VAE: sdxl-vae to bfloat safetensor=false before cuda then attn_proc / scale factor 8 \n")
         f.write(f"Model UNET: realistic_porn \n")
-        f.write(f"Model HiDiffusion OFF \n")
-        f.write(f"Model do_resize ON \n")
-        f.write(f"added torch to prereq and changed accellerate  \n")
     upload_to_ftp(filename) 
     
 @spaces.GPU(duration=40)
@@ -252,7 +271,7 @@ def generate_30(
     uploadNote(prompt,num_inference_steps,guidance_scale,timestamp)
     batch_options = options.copy()
     rv_image = pipe(**batch_options).images[0]
-    sd_image_path = f"rv50_A_{timestamp}.png"
+    sd_image_path = f"rv50_E_{timestamp}.png"
     rv_image.save(sd_image_path,optimize=False,compress_level=0)
     upload_to_ftp(sd_image_path)
     image_paths = save_image(rv_image)
@@ -262,7 +281,7 @@ def generate_30(
     with torch.no_grad():
         upscale = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
     downscale1 = upscale.resize((upscale.width // 4, upscale.height // 4), Image.LANCZOS)
-    downscale_path = f"rv50_upscale_{timestamp}.png"
+    downscale_path = f"rv50E_upscale_{timestamp}.png"
     downscale1.save(downscale_path,optimize=False,compress_level=0)
     upload_to_ftp(downscale_path) 
     image_paths = [save_image(downscale1)]
@@ -306,7 +325,7 @@ def generate_60(
     uploadNote(prompt,num_inference_steps,guidance_scale,timestamp)
     batch_options = options.copy()
     rv_image = pipe(**batch_options).images[0]
-    sd_image_path = f"rv50_A_{timestamp}.png"
+    sd_image_path = f"rv50_E_{timestamp}.png"
     rv_image.save(sd_image_path,optimize=False,compress_level=0)
     upload_to_ftp(sd_image_path)
     image_paths = save_image(rv_image)
@@ -314,7 +333,7 @@ def generate_60(
     with torch.no_grad():
         upscale = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
     downscale1 = upscale.resize((upscale.width // 4, upscale.height // 4), Image.LANCZOS)
-    downscale_path = f"rv50_upscale_{timestamp}.png"
+    downscale_path = f"rv50E_upscale_{timestamp}.png"
     downscale1.save(downscale_path,optimize=False,compress_level=0)
     upload_to_ftp(downscale_path) 
     image_paths = [save_image(downscale1)]
@@ -358,7 +377,7 @@ def generate_90(
     uploadNote(prompt,num_inference_steps,guidance_scale,timestamp)
     batch_options = options.copy()
     rv_image = pipe(**batch_options).images[0]
-    sd_image_path = f"rv50_A_{timestamp}.png"
+    sd_image_path = f"rv50_E_{timestamp}.png"
     rv_image.save(sd_image_path,optimize=False,compress_level=0)
     upload_to_ftp(sd_image_path)
     image_paths = save_image(rv_image)
@@ -366,7 +385,7 @@ def generate_90(
     with torch.no_grad():
         upscale = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
     downscale1 = upscale.resize((upscale.width // 4, upscale.height // 4), Image.LANCZOS)
-    downscale_path = f"rv50_upscale_{timestamp}.png"
+    downscale_path = f"rv50E_upscale_{timestamp}.png"
     downscale1.save(downscale_path,optimize=False,compress_level=0)
     upload_to_ftp(downscale_path) 
     image_paths = [save_image(downscale1)]