Update app.py

app.py

@@ -25,7 +25,6 @@ import datetime
 from gradio import themes
 from hidiffusion import apply_hidiffusion, remove_hidiffusion
 import gc
-from diffusers.utils.torch_utils import randn_tensor
 
 torch.backends.cuda.matmul.allow_tf32 = False
 torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
@@ -149,7 +148,7 @@ def load_and_prepare_model(model_id):
     #pipe.scheduler=EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config, beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012, steps_offset=1)
     #pipe.scheduler=EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear")
 
-    pipe.to(device=torch.device('cuda'), dtype=torch.bfloat16)
+    pipe.to(device=device, dtype=torch.bfloat16)
     #pipe.to(torch.bfloat16)
     
     #apply_hidiffusion(pipe)
@@ -252,9 +251,7 @@ def generate_30(
     #if juggernaut == True:
     #    pipe.vae=vaeX
     seed = int(randomize_seed_fn(seed, randomize_seed))
-    generator = torch.Generator(device=torch.device('cuda')).manual_seed(seed)
-    latent_size = int(height / 8)
-    latent_input = randn_tensor(shape=(1,4,latent_size,latent_size),generator=generator, device=torch.device('cuda'), dtype=torch.bfloat16)
+    generator = torch.Generator(device='cuda').manual_seed(seed)
     #prompt, negative_prompt = apply_style(style_selection, prompt, negative_prompt)
     options = {
         "prompt": [prompt],
@@ -268,12 +265,11 @@ def generate_30(
         "generator": generator,
      #   "timesteps": sampling_schedule,
         "output_type": "pil",
-        "latents": latent_input
     }
     if use_resolution_binning:
         options["use_resolution_binning"] = True
     images = []
-    pipe.scheduler.set_timesteps(num_inference_steps,torch.device('cuda'))
+    pipe.scheduler.set_timesteps(num_inference_steps,device)
     timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
     uploadNote(prompt,num_inference_steps,guidance_scale,timestamp,denoise)
     batch_options = options.copy()
@@ -306,7 +302,7 @@ def generate_60(
     denoise: float = 0.3,
     progress=gr.Progress(track_tqdm=True)  # Add progress as a keyword argument
 ):
-    torch.backends.cudnn.benchmark = False
+    torch.backends.cudnn.benchmark = True
     torch.cuda.empty_cache()
     gc.collect()
     global models
@@ -314,9 +310,7 @@ def generate_60(
     #if juggernaut == True:
     #    pipe.vae=vaeX
     seed = int(randomize_seed_fn(seed, randomize_seed))
-    generator = torch.Generator(device=torch.device('cuda')).manual_seed(seed)
-    latent_size = int(height / 8)
-    latent_input = randn_tensor(shape=(1,4,latent_size,latent_size),generator=generator, device=torch.device('cuda'), dtype=torch.bfloat16)
+    generator = torch.Generator(device='cuda').manual_seed(seed)
     #prompt, negative_prompt = apply_style(style_selection, prompt, negative_prompt)
     options = {
         "prompt": [prompt],
@@ -330,12 +324,11 @@ def generate_60(
         "generator": generator,
      #   "timesteps": sampling_schedule,
         "output_type": "pil",
-        "latents": latent_input
     }
     if use_resolution_binning:
         options["use_resolution_binning"] = True
     images = []
-    pipe.scheduler.set_timesteps(num_inference_steps,torch.device('cuda'))
+    pipe.scheduler.set_timesteps(num_inference_steps,device)
     timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
     uploadNote(prompt,num_inference_steps,guidance_scale,timestamp,denoise)
     batch_options = options.copy()
@@ -368,7 +361,7 @@ def generate_90(
     denoise: float = 0.3,
     progress=gr.Progress(track_tqdm=True)  # Add progress as a keyword argument
 ):
-    torch.backends.cudnn.benchmark = False
+    torch.backends.cudnn.benchmark = True
     torch.cuda.empty_cache()
     gc.collect()
     global models
@@ -376,9 +369,7 @@ def generate_90(
     #if juggernaut == True:
     #    pipe.vae=vaeX
     seed = int(randomize_seed_fn(seed, randomize_seed))
-    generator = torch.Generator(device=torch.device('cuda')).manual_seed(seed)
-    latent_size = int(height / 8)
-    latent_input = randn_tensor(shape=(1,4,latent_size,latent_size),generator=generator, device=torch.device('cuda'), dtype=torch.bfloat16)
+    generator = torch.Generator(device='cuda').manual_seed(seed)
     #prompt, negative_prompt = apply_style(style_selection, prompt, negative_prompt)
     options = {
         "prompt": [prompt],
@@ -392,12 +383,11 @@ def generate_90(
         "generator": generator,
      #   "timesteps": sampling_schedule,
         "output_type": "pil",
-        "latents": latent_input
     }
     if use_resolution_binning:
         options["use_resolution_binning"] = True
     images = []
-    pipe.scheduler.set_timesteps(num_inference_steps,torch.device('cuda'))
+    pipe.scheduler.set_timesteps(num_inference_steps,device)
     timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
     uploadNote(prompt,num_inference_steps,guidance_scale,timestamp,denoise)
     batch_options = options.copy()