fix: VARCO 모델 사전 다운로드로 콜드 스타트 타임아웃 해결

app.py

@@ -306,6 +306,47 @@ class ModelManager:
 # Global model manager
 model_manager = ModelManager()
 
+# ============================================================
+# 베이스 모델 사전 캐싱 (콜드 스타트 방지)
+# ============================================================
+
+def preload_base_models():
+    """Pre-download base models to avoid cold start timeout"""
+    if not TORCH_AVAILABLE:
+        print("Skipping preload: PyTorch not available")
+        return
+
+    from huggingface_hub import snapshot_download
+    import os
+
+    # Models that need pre-caching (large or slow to download)
+    models_to_cache = [
+        "NCSOFT/Llama-VARCO-8B-Instruct",  # VARCO - 16GB, often times out on first load
+    ]
+
+    print("=" * 50)
+    print("Pre-downloading base models (this may take a while)...")
+    print("=" * 50)
+
+    for model_id in models_to_cache:
+        try:
+            print(f"  Downloading: {model_id}")
+            # Download all model files to HF cache
+            cache_dir = snapshot_download(
+                repo_id=model_id,
+                ignore_patterns=["*.md", "*.txt"],  # Skip docs
+            )
+            print(f"    ✓ Downloaded to: {cache_dir}")
+
+        except Exception as e:
+            print(f"    ✗ Failed to download {model_id}: {e}")
+
+    print("Pre-download complete!")
+    print("=" * 50)
+
+# Run preload at startup
+preload_base_models()
+
 # ============================================================
 # 시스템 프롬프트 생성
 # ============================================================