handler.py

# handler.py

from transformers import pipeline, AutoModelForSeq2SeqLM, AutoTokenizer

MODEL_ID = "Salesforce/codet5-base-multi-sum"

# Load once at startup
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model     = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)
summarizer = pipeline(
    "text2text-generation",
    model=model,
    tokenizer=tokenizer,
    device=-1  # CPU
)

def inference(inputs: dict):
    """
    Hugging Face will call this for every request.
    `inputs` is like {"inputs": "<the diff patch text>"}
    """
    patch = inputs.get("inputs", "")
    out = summarizer(
        patch,
        max_length=64,
        truncation=True,
        num_beams=4,
        early_stopping=True
    )
    # Return in the HF expected format (list of dicts)
    return [{"summary_text": out[0]["generated_text"]}]