torch_compile/run_vit.py

#!/usr/bin/env python3
# Vision Transformer (ViT) image-classification on Neuron – full graph, constant shapes
import argparse
import logging
import time
import torch
from transformers import AutoImageProcessor, ViTForImageClassification
from datasets import load_dataset
import torch_neuronx  # guarantees Neuron backend

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


def main():
    parser = argparse.ArgumentParser(description="ViT on Neuron (full graph)")
    parser.add_argument("--model", default="google/vit-base-patch16-224")
    args = parser.parse_args()

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

    # load dataset image
    dataset = load_dataset("huggingface/cats-image")
    image = dataset["test"]["image"][0]

    # load processor & model
    processor = AutoImageProcessor.from_pretrained(args.model)
    model = ViTForImageClassification.from_pretrained(
        args.model, torch_dtype=torch.float32, attn_implementation="eager"
    ).eval()

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

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

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

    # warmup
    warmup_start = time.time()
    with torch.no_grad():
        _ = model(**inputs)
    logger.info("Warmup: %.3f s", 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 ImageNet class
    predicted_class_idx = logits.argmax(-1).item()
    predicted_label = model.config.id2label[predicted_class_idx]

    logger.info("Run: %.3f s", run_time)
    logger.info("Predicted label: %s", predicted_label)


if __name__ == "__main__":
    main()