import spaces import gradio as gr from transformers import ( AutoTokenizer, AutoModelForSequenceClassification, TextClassificationPipeline, ) MODEL_REPO = "Ahmed-AI-Engineer/smartassist-nlp-intent-model" CONFIDENCE_THRESHOLD = 0.60 print(f"Loading model from: {MODEL_REPO}") # ----------------------------------- # Load tokenizer # ----------------------------------- tokenizer = AutoTokenizer.from_pretrained( MODEL_REPO ) # ----------------------------------- # Load model # ----------------------------------- model = AutoModelForSequenceClassification.from_pretrained( MODEL_REPO ) # IMPORTANT: # ZeroGPU supports CUDA placement during startup # through its CUDA emulation layer. model = model.to("cuda") model.eval() print("Model loaded successfully.") # ----------------------------------- # Create pipeline # ----------------------------------- classifier = TextClassificationPipeline( model=model, tokenizer=tokenizer, device=0 ) # ----------------------------------- # Prediction function # ----------------------------------- @spaces.GPU(duration=30) def predict_intent(text: str): text = text.strip() if not text: return { "intent": "", "confidence": 0.0, "needs_confirmation": True, "raw_text": text, } result = classifier(text)[0] confidence = float(result["score"]) return { "intent": result["label"], "confidence": confidence, "needs_confirmation": confidence < CONFIDENCE_THRESHOLD, "raw_text": text, } # ----------------------------------- # Gradio UI # ----------------------------------- with gr.Blocks( title="SmartAssist NLP API" ) as demo: gr.Markdown( """ # SmartAssist NLP API Intent classification for SmartAssist home automation commands. Supports: - English - Urdu - Roman Urdu """ ) text_input = gr.Textbox( label="Voice command text", placeholder="Example: light on karo" ) predict_button = gr.Button( "Predict", variant="primary" ) prediction_output = gr.JSON( label="Prediction" ) predict_button.click( fn=predict_intent, inputs=text_input, outputs=prediction_output, api_name="predict" ) text_input.submit( fn=predict_intent, inputs=text_input, outputs=prediction_output, api_name=False ) # ----------------------------------- # Launch Space # ----------------------------------- if __name__ == "__main__": demo.launch()