torch_compile/run_albert.py

import argparse
import logging
import time

import torch
from transformers import AutoTokenizer, AlbertForSequenceClassification

import torch_neuronx  # ensure Neuron backend is available

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


def main():
    parser = argparse.ArgumentParser(description="Run ALBERT on Neuron")
    parser.add_argument(
        "--model", type=str, default="albert-base-v2", help="ALBERT model name"
    )
    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 ALBERT model and tokenizer
    model = AlbertForSequenceClassification.from_pretrained(
        args.model, torch_dtype=torch.float32, attn_implementation="eager"
    )
    model.eval()

    tokenizer = AutoTokenizer.from_pretrained(args.model)
    inputs = tokenizer(
        "Hamilton is considered to be the best musical of human history.",
        return_tensors="pt"
    )

    # Pre-run once to fix shapes before compilation
    with torch.no_grad():
        _ = model(**inputs).logits

    # Compile forward pass
    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

    # Actual run
    run_start = time.time()
    with torch.no_grad():
        logits = model(**inputs).logits
    run_time = time.time() - run_start
    predicted_class_id = logits.argmax().item()
    predicted_class_label = model.config.id2label[predicted_class_id]

    logger.info(f"Warmup: {warmup_time:.2f}s, Run: {run_time:.4f}s")
    logger.info(f"Output label: {predicted_class_label}")


if __name__ == "__main__":
    main()