Update app.py

app.py

@@ -2,7 +2,6 @@ import spaces
 import gradio as gr
 import os
 import sys
-import argparse
 import random
 import time
 from omegaconf import OmegaConf
@@ -31,23 +30,19 @@ def download_model():
         if not os.path.exists(local_file):
             hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir='./checkpoints/dynamicrafter_1024_v1/', force_download=True)
 
-
-
 download_model()
 ckpt_path='checkpoints/dynamicrafter_1024_v1/model.ckpt'
 config_file='configs/inference_1024_v1.0.yaml'
 config = OmegaConf.load(config_file)
 model_config = config.pop("model", OmegaConf.create())
-model_config['params']['unet_config']['params']['use_checkpoint']=False   
+model_config['params']['unet_config']['params']['use_checkpoint']=True  # Checkpoint 사용하여 메모리 사용 최적화
 model = instantiate_from_config(model_config)
 assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!"
 model = load_model_checkpoint(model, ckpt_path)
 model.eval()
 model = model.cuda()
 
-
-
-@spaces.GPU(duration=300)
+@spaces.GPU(duration=300, gpu_type="h100")  # H100 GPU 사용 지정
 def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
     resolution = (576, 1024)
     save_fps = 8
@@ -56,31 +51,29 @@ def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
         transforms.Resize(min(resolution)),
         transforms.CenterCrop(resolution),
         ])
-    torch.cuda.empty_cache()
-    print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time())))
+    torch.cuda.empty_cache()  # GPU 캐시 메모리 정리
+    print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
     start = time.time()
     if steps > 60:
         steps = 60 
 
-    batch_size=1
+    batch_size = 1
     channels = model.model.diffusion_model.out_channels
     frames = model.temporal_length
     h, w = resolution[0] // 8, resolution[1] // 8
     noise_shape = [batch_size, channels, frames, h, w]
 
-    # text cond
-    with torch.no_grad(), torch.cuda.amp.autocast():
+    # 텍스트 조건 생성
+    with torch.no_grad(), torch.cuda.amp.autocast():  # 메모리 사용량 감소 및 연산 속도 개선
         text_emb = model.get_learned_conditioning([prompt])
-    
-        # img cond
+        
+        # 이미지 조건 생성
         img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device)
         img_tensor = (img_tensor / 255. - 0.5) * 2
-    
         image_tensor_resized = transform(img_tensor) #3,256,256
         videos = image_tensor_resized.unsqueeze(0) # bchw
         
         z = get_latent_z(model, videos.unsqueeze(2)) #bc,1,hw
-        
         img_tensor_repeat = repeat(z, 'b c t h w -> b c (repeat t) h w', repeat=frames)
     
         cond_images = model.embedder(img_tensor.unsqueeze(0)) ## blc
@@ -91,9 +84,9 @@ def infer(image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
         fs = torch.tensor([fs], dtype=torch.long, device=model.device)
         cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat]}
         
-        ## inference
+        # 추론
         batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale)
-        ## b,samples,c,t,h,w
+        # b,samples,c,t,h,w
     
         video_path = './output.mp4'
         save_videos(batch_samples, './', filenames=['output'], fps=save_fps)