fix load gpu

app.py

@@ -174,7 +174,10 @@ def patch_paths_in_exp_config(exp_cfg: Dict[str, Any], repo_root: Path, qwen_roo
 # 学长说的就是这里：所有费资源的操作（加载+推理）都要放在这里面
 @spaces.GPU
 def run_edit(audio_file, caption, num_steps, guidance_scale, guidance_rescale, seed):
-    # 延迟导入，防止全局污染
+    
+
+
+
     import torch
     import hydra
     from omegaconf import OmegaConf
@@ -230,8 +233,18 @@ def run_edit(audio_file, caption, num_steps, guidance_scale, guidance_rescale, s
         logger.info("Moving model to CUDA (FP16)...")
         
         # 这一步将模型送入显卡
-        model = model.to(device, dtype=torch.float16).eval()
-        
+        def safe_move_model(m, dev):
+            logger.info("🛡️ Moving model layer by layer to avoid RAM spike...")
+            for name, child in m.named_children():
+                # 逐层搬运：CPU内存释放一点 -> GPU显存增加一点
+                child.to(dev, dtype=torch.float16)
+                logger.info(f"Moving {name} to GPU...")
+            m.to(dev, dtype=torch.float16)
+            return m
+            
+        model = safe_move_model(model, device)
+        model.eval()
+        logger.info("Model is moved to CUDA.")
         # Scheduler
         try:
             scheduler = noise_schedulers.DDIMScheduler.from_pretrained(