Update app.py

app.py

@@ -51,21 +51,28 @@ EXAMPLES = [
 ]
 
 # --- On ZeroGPU, torch is patched at import time and there is NO GPU at
-#     module scope, so we cannot load model weights here (even a CPU load of
-#     safetensors gets intercepted and tries to reach CUDA). We load only the
-#     tokenizer eagerly, and defer ALL model loading to the first GPU call. ---
-print("Loading tokenizer...")
+#     module scope, so we cannot load model weights onto CUDA here. But we CAN
+#     pre-download the files to the local cache at startup, so the GPU call
+#     only has to load from disk (fast) and stays within the 60s GPU budget. ---
+from huggingface_hub import snapshot_download
+
+print("Loading tokenizer and pre-downloading model files...")
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
 tokenizer.pad_token = tokenizer.eos_token
-print("Tokenizer ready; model loads on first request.")
+
+# Warm the HF cache so weights are on local disk before any GPU call.
+snapshot_download(BASE_MODEL)
+snapshot_download(ADAPTER_ID, token=HF_TOKEN)
+print("Tokenizer ready and weights cached; model loads on first request.")
 
 _model = None  # lazily populated inside the @spaces.GPU function
 
 
 def _load_model():
     """
-    Load base + LoRA adapter and place on CUDA. Called once, lazily, from
-    inside the @spaces.GPU function where a GPU is actually attached.
+    Load base + LoRA adapter from the local cache and place on CUDA. Called
+    once, lazily, from inside the @spaces.GPU function where a GPU is
+    attached. Because files are already cached, this is just a disk load.
     """
     global _model
     if _model is not None:
@@ -100,7 +107,7 @@ def _decode(output_ids, input_len: int) -> str:
     return tokenizer.decode(gen, skip_special_tokens=True).strip()
 
 
-@spaces.GPU(duration=300)
+@spaces.GPU(duration=60)
 def rewrite(text: str):
     """
     Produce base ('before') and fine-tuned ('after') rewrites.