load in advance

app.py

@@ -8,16 +8,18 @@ from free_lunch_utils import register_free_upblock2d, register_free_crossattn_up
 
 from pipeline_freescale import StableDiffusionXLPipeline
 from pipeline_freescale_turbo import StableDiffusionXLPipeline_Turbo
+
+dtype = torch.bfloat16
+device = "cuda" if torch.cuda.is_available() else "cpu"
 model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
 model_ckpt_turbo = "stabilityai/sdxl-turbo"
+pipe = StableDiffusionXLPipeline.from_pretrained(model_ckpt, torch_dtype=dtype).to(device)
+pipe_turbo = StableDiffusionXLPipeline_Turbo.from_pretrained(model_ckpt_turbo, torch_dtype=dtype).to(device)
+generator = torch.Generator(device='cuda')
+torch.cuda.empty_cache()
 
-pipe = StableDiffusionXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
-pipe_turbo = StableDiffusionXLPipeline_Turbo.from_pretrained(model_ckpt_turbo, torch_dtype=torch.float16)
-
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=100)
 def infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu, restart_steps):
-    pipe = pipe.to("cuda") 
-    generator = torch.Generator(device='cuda')
     generator = generator.manual_seed(seed)
     if not disable_freeu:
         register_free_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)
@@ -29,10 +31,8 @@ def infer_gpu_part(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_sca
                 ).images[0]
     return result
 
-@spaces.GPU(duration=40)
+@spaces.GPU(duration=20)
 def infer_gpu_part_turbo(pipe, seed, prompt, negative_prompt, ddim_steps, guidance_scale, resolutions_list, fast_mode, cosine_scale, disable_freeu, restart_steps):
-    pipe = pipe.to("cuda") 
-    generator = torch.Generator(device='cuda')
     generator = generator.manual_seed(seed)
     if not disable_freeu:
         register_free_upblock2d(pipe, b1=1.1, b2=1.2, s1=0.6, s2=0.4)