Update app.py

app.py

@@ -1,15 +1,15 @@
 import spaces
 import os
 
-os.putenv('PYTORCH_NVML_BASED_CUDA_CHECK','1')
-os.putenv('TORCH_LINALG_PREFER_CUSOLVER','1')
+ # os.putenv('PYTORCH_NVML_BASED_CUDA_CHECK','1')
+# os.putenv('TORCH_LINALG_PREFER_CUSOLVER','1')
 alloc_conf_parts = [
     'expandable_segments:True',
     'pinned_use_background_threads:True'  # Specific to pinned memory.
 ]
-os.environ['PYTORCH_CUDA_ALLOC_CONF'] = ','.join(alloc_conf_parts)
-os.environ["SAFETENSORS_FAST_GPU"] = "1"
-os.putenv('HF_HUB_ENABLE_HF_TRANSFER','1')
+# os.environ['PYTORCH_CUDA_ALLOC_CONF'] = ','.join(alloc_conf_parts)
+# os.environ["SAFETENSORS_FAST_GPU"] = "1"
+# os.putenv('HF_HUB_ENABLE_HF_TRANSFER','1')
 
 import gradio as gr
 import numpy as np
@@ -27,14 +27,14 @@ import time
 from image_gen_aux import UpscaleWithModel
 from huggingface_hub import hf_hub_download
 import datetime
-import cyper
+#import cyper
 
 from diffusers import AutoencoderKL
 #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
@@ -44,10 +44,10 @@ 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'''
+# code = r'''
 import torch
 import paramiko
 import os
@@ -68,9 +68,9 @@ def upload_to_ftp(filename):
         print(f"Uploaded {filename} to FTP server")
     except Exception as e:
         print(f"FTP upload error: {e}")
-'''
+# '''
 
-pyx = cyper.inline(code, fast_indexing=True, directives=dict(boundscheck=False, wraparound=False, language_level=3))
+# 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)
@@ -82,11 +82,11 @@ pipe = StableDiffusion3Pipeline.from_pretrained(
     #vae=None,
     #vae=AutoencoderKL.from_pretrained("ford442/stable-diffusion-3.5-large-fp32", use_safetensors=True, subfolder='vae',token=True),
      #scheduler = FlowMatchHeunDiscreteScheduler.from_pretrained('ford442/stable-diffusion-3.5-large-bf16', subfolder='scheduler',token=True),
-    text_encoder=None, #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True),
+#    text_encoder=None, #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True),
     # text_encoder=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True),
-    text_encoder_2=None, #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True),
+#    text_encoder_2=None, #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True),
   #  text_encoder_2=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True),
-    text_encoder_3=None, #T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True),
+#    text_encoder_3=None, #T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True),
   #  text_encoder_3=T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True),
     #tokenizer=CLIPTokenizer.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", add_prefix_space=True, subfolder="tokenizer", token=True),
     #tokenizer_2=CLIPTokenizer.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", add_prefix_space=True, subfolder="tokenizer_2", token=True),
@@ -95,9 +95,9 @@ pipe = StableDiffusion3Pipeline.from_pretrained(
     #torch_dtype=torch.bfloat16,
     #use_safetensors=False,
 )
-text_encoder=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True).to(torch.device("cuda:0")) #, dtype=torch.bfloat16)
-text_encoder_2=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True).to(torch.device("cuda:0")) #, dtype=torch.bfloat16)
-text_encoder_3=T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-fp32", subfolder='text_encoder_3',token=True).to(torch.device("cuda:0")) #, dtype=torch.bfloat16)
+#text_encoder=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True).to(torch.device("cuda:0")) #, dtype=torch.bfloat16)
+#text_encoder_2=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True).to(torch.device("cuda:0")) #, dtype=torch.bfloat16)
+#text_encoder_3=T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-fp32", subfolder='text_encoder_3',token=True).to(torch.device("cuda:0")) #, dtype=torch.bfloat16)
     
 pipe.load_lora_weights("ford442/sdxl-vae-bf16", weight_name="LoRA/UltraReal.safetensors")
 
@@ -122,9 +122,9 @@ def infer_30(
     num_inference_steps,
     progress=gr.Progress(track_tqdm=True),
 ):
-    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)
+    #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 --')
@@ -147,7 +147,8 @@ def infer_30(
     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)
+    #pyx.upload_to_ftp(sd35_path)
+    upload_to_ftp(sd35_path)
     #  pipe.unet.to('cpu')
     upscaler_2.to(torch.device('cuda'))
     with torch.no_grad():
@@ -156,7 +157,8 @@ def infer_30(
     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)
+    #pyx.upload_to_ftp(upscale_path)
+    upload_to_ftp(upscale_path)
     return sd_image, prompt
 
 @spaces.GPU(duration=70)
@@ -171,9 +173,9 @@ def infer_60(
     num_inference_steps,
     progress=gr.Progress(track_tqdm=True),
 ):
-    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)
+    #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 --')
@@ -196,16 +198,18 @@ def infer_60(
     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)
+    #pyx.upload_to_ftp(sd35_path)
+    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 // 2, upscale2.height // 2),Image.LANCZOS)
+    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)
+    #pyx.upload_to_ftp(upscale_path)
+    upload_to_ftp(upscale_path)
     return sd_image, prompt
 
 @spaces.GPU(duration=100)
@@ -220,9 +224,9 @@ def infer_90(
     num_inference_steps,
     progress=gr.Progress(track_tqdm=True),
 ):
-    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)
+    #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 --')
@@ -245,16 +249,18 @@ def infer_90(
     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)
+    #pyx.upload_to_ftp(sd35_path)
+    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 // 2, upscale2.height // 2),Image.LANCZOS)
+    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)
+    #pyx.upload_to_ftp(upscale_path)
+    upload_to_ftp(upscale_path)
     return sd_image, prompt
 
 @spaces.GPU(duration=110)
@@ -269,6 +275,9 @@ def infer_100(
     num_inference_steps,
     progress=gr.Progress(track_tqdm=True),
 ):
+    #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 --')
@@ -283,6 +292,7 @@ def infer_100(
             num_inference_steps=num_inference_steps,
             width=width,
             height=height,
+         #   cross_attention_kwargs={"scale": 0.75},
             generator=generator,
             max_sequence_length=512
     ).images[0]
@@ -290,16 +300,18 @@ def infer_100(
     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)
+    #pyx.upload_to_ftp(sd35_path)
+    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 // 2, upscale2.height // 2),Image.LANCZOS)
+    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)
+    #pyx.upload_to_ftp(upscale_path)
+    upload_to_ftp(upscale_path)
     return sd_image, prompt
     
 css = """