torch_compile/run_sam2.py

#!/usr/bin/env python3
# SAM encoder on Neuron – constant-shape, no lambda
import argparse
import logging
import time
import torch
from transformers import SamProcessor, SamModel
from PIL import Image
import torch_neuronx  # guarantees Neuron backend

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


def main():
    parser = argparse.ArgumentParser(description="SAM encoder on Neuron (full graph)")
    parser.add_argument("--model", default="facebook/sam-vit-base")
    args = parser.parse_args()

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

    # load processor & model
    processor = SamProcessor.from_pretrained(args.model)
    model = SamModel.from_pretrained(args.model, attn_implementation="eager").eval()

    # dummy 224×224 RGB image
    dummy_image = Image.new("RGB", (224, 224), color="red")
    # constant-shape inputs (no points → encoder only)
    inputs = processor(images=dummy_image, return_tensors="pt")

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

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

    # warmup
    start = time.time()
    with torch.no_grad():
        _ = model.get_image_embeddings(**inputs)
    logger.info("Warmup: %.3f s", time.time() - start)

    # benchmark
    start = time.time()
    with torch.no_grad():
        embeddings = model.get_image_embeddings(**inputs)
    logger.info("Run: %.3f s", time.time() - start)
    logger.info("Embedding shape: %s", embeddings.shape)  # [1, 256, 64, 64]


if __name__ == "__main__":
    main()