Update app.py

app.py

@@ -13,7 +13,6 @@ from peft import PeftModel
 BASE_MODEL = "unsloth/gemma-3-270m-it"
 LORA_PATH  = "Razon2006/gemma3-tamil-Grapheme-NewChange-r32-lr"
 
-# Retrieve the Hugging Face token from environment secrets
 hf_token = os.environ.get("HF_TOKEN")
 
 print("Loading tokenizer...")
@@ -27,12 +26,11 @@ base_model = AutoModelForCausalLM.from_pretrained(
 )
 
 print("Loading LoRA adapter...")
-# Load the model globally and move to CUDA. 
-# ZeroGPU safely intercepts standard PyTorch .to("cuda") calls.
 model = PeftModel.from_pretrained(base_model, LORA_PATH, token=hf_token)
-model.to("cuda")
-model.eval()
-print("Model ready!")
+
+# Best practice for ZeroGPU: Keep model in CPU RAM globally.
+# We will move it to the GPU inside the annotated function.
+print("Model loaded into memory!")
 
 # ─── Grapheme Tokenizer & Constants ───────────────────────────────────────────
 INSTRUCTION = (
@@ -89,6 +87,9 @@ def post_process_grapheme(text):
 # ─── Inference Pipeline ───────────────────────────────────────────────────────
 @spaces.GPU(duration=60)
 def generate_pathaseth(venba, max_new_tokens=256):
+    # Explicitly move model to GPU here. ZeroGPU safely grants hardware access at this exact moment.
+    model.to("cuda")
+
     venba = venba.strip()
     if not venba:
         return "Please enter a Venba verse."