Remove CUDA environment variable setting and enhance GPU handling in run_inference function

app.py

@@ -25,7 +25,6 @@ CHECKPOINT_EXPECTED = [
 ]
 
 os.makedirs(OUTPUT_ROOT, exist_ok=True)
-os.environ.setdefault("CUDA_VISIBLE_DEVICES", str(DEFAULT_CONFIG.get("GPU_id", "0")))
 
 
 def ensure_checkpoints() -> None:
@@ -138,9 +137,7 @@ def run_inference(
     musical_mask_start: float,
     musical_mask_end: float,
     seed: Optional[float],
-):
-    if not torch.cuda.is_available():
-        raise gr.Error("This Space has no GPU attached. Please run locally with a GPU or duplicate to a GPU Space.")
+):    
 
     condition_type = _validate_condition_choices(condition_type)
     config = _build_base_config()
@@ -181,7 +178,8 @@ def run_inference(
 
     try:
         models = model_cache.get(config)
-        pipe = models["pipe"]
+        pipe = models["pipe"].to("cuda")
+        pipe.enable_attention_slicing()
         pipe.scheduler.config.sigma_min = config["sigma_min"]
         pipe.scheduler.config.sigma_max = config["sigma_max"]
         prompt_for_model = "" if config["no_text"] else (prompt_text or "")