torch_compile/run_mobilebert.py

#!/usr/bin/env python3
# MobileBERT text-classification on Neuron
import argparse
import logging
import time

import torch
from transformers import AutoTokenizer, MobileBertForSequenceClassification
import torch_neuronx  # ensures Neuron backend

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)


def main():
    parser = argparse.ArgumentParser(description="Run MobileBERT on Neuron")
    parser.add_argument(
        "--model",
        type=str,
        default="google/mobilebert-uncased",
        help="MobileBERT model name on Hugging Face Hub",
    )
    parser.add_argument("--batch-size", type=int, default=1, help="Batch size")
    args = parser.parse_args()

    torch.set_default_dtype(torch.float32)
    torch.manual_seed(42)

    # load tokenizer & model
    tokenizer = AutoTokenizer.from_pretrained(args.model)
    model = MobileBertForSequenceClassification.from_pretrained(
        args.model, torch_dtype=torch.float32, attn_implementation="eager"
    ).eval()

    # tokenize sample
    text = "MobileBERT is a compact BERT for on-device NLP."
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)

    # pre-run to lock shapes
    with torch.no_grad():
        _ = model(**inputs).logits

    # compile
    model.forward = torch.compile(model.forward, backend="neuron", fullgraph=True)

    # warmup
    warmup_start = time.time()
    with torch.no_grad():
        _ = model(**inputs)
    warmup_time = time.time() - warmup_start

    # benchmark run
    run_start = time.time()
    with torch.no_grad():
        logits = model(**inputs).logits
    run_time = time.time() - run_start

    # top-1 label
    predicted_class_id = logits.argmax().item()
    predicted_label = model.config.id2label[predicted_class_id]

    logger.info("Warmup: %.2f s, Run: %.4f s", warmup_time, run_time)
    logger.info("Predicted label: %s", predicted_label)


if __name__ == "__main__":
    main()