app.py

import gradio as gr
import llm_blender

# Load the model and explicitly force it to use the CPU
blender = llm_blender.Blender()
blender.loadranker("llm-blender/PairRM", device="cpu")

def score_pair(prompt, cand_a, cand_b):
    inputs = [prompt]
    candidates_A = [cand_a]
    candidates_B = [cand_b]
    
    # Compare returns a list of booleans (True if A is better)
    comparison_results = blender.compare(inputs, candidates_A, candidates_B)
    
    return "A" if comparison_results[0] else "B"

# Create the API endpoint
iface = gr.Interface(
    fn=score_pair,
    inputs=["text", "text", "text"],
    outputs="text"
)

iface.launch()