torch_compile/run_convnext.py

import argparse
import logging
import time

import torch
from transformers import AutoImageProcessor, ConvNextForImageClassification
from datasets import load_dataset
import torch_neuronx  # ensures Neuron backend is available

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


def main():
    parser = argparse.ArgumentParser(
        description="ConvNeXt image-classification with torch.compile on Neuron"
    )
    parser.add_argument(
        "--model",
        type=str,
        default="facebook/convnext-tiny-224",
        help="ConvNeXT 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 dataset and pick an image
    dataset = load_dataset("huggingface/cats-image")
    image = dataset["test"]["image"][0]

    # Load processor and model
    processor = AutoImageProcessor.from_pretrained(args.model)
    model = ConvNextForImageClassification.from_pretrained(
        args.model, torch_dtype=torch.float32, attn_implementation="eager"
    )
    model.eval()

    # Preprocess image
    inputs = processor(images=image, return_tensors="pt")

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

    # Compile forward pass (allow graph breaks to avoid instruction-limit)
    model.forward = torch.compile(model.forward, backend="neuron", fullgraph=False)

    # 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():
        outputs = model(**inputs)
    run_time = time.time() - run_start

    # Predicted ImageNet class
    predicted_class_idx = outputs.logits.argmax(-1).item()
    predicted_label = model.config.id2label[predicted_class_idx]

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


if __name__ == "__main__":
    main()