LightGlueProject/LightGlue_Deployment/scripts/build_tensorrt.py

#!/usr/bin/env python3
"""
Convert exported ONNX graphs into TensorRT engines.

Requires a local TensorRT installation with the Python bindings available.
"""

from __future__ import annotations

import argparse
from pathlib import Path

import tensorrt as trt


def build_engine(onnx_path: Path, engine_path: Path, fp16: bool = True, workspace_gb: int = 2):
    logger = trt.Logger(trt.Logger.WARNING)
    builder = trt.Builder(logger)
    network_flags = trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH
    network = builder.create_network(network_flags)
    parser = trt.OnnxParser(network, logger)

    onnx_bytes = onnx_path.read_bytes()
    if not parser.parse(onnx_bytes):
        for idx in range(parser.num_errors):
            print(parser.get_error(idx))
        raise RuntimeError(f"Failed to parse ONNX: {onnx_path}")

    config = builder.create_builder_config()
    config.max_workspace_size = workspace_gb * (1 << 30)
    if fp16 and builder.platform_has_fast_fp16:
        config.set_flag(trt.BuilderFlag.FP16)

    profile = builder.create_optimization_profile()
    for idx in range(network.num_inputs):
        tensor = network.get_input(idx)
        shape = tuple(tensor.shape)
        profile.set_shape(tensor.name, shape, shape, shape)
    config.add_optimization_profile(profile)

    engine = builder.build_engine(network, config)
    if engine is None:
        raise RuntimeError(f"Failed to build TensorRT engine for {onnx_path}")

    engine_path.parent.mkdir(parents=True, exist_ok=True)
    with engine_path.open("wb") as f:
        f.write(engine.serialize())

    print(f"[OK] TensorRT engine saved to: {engine_path}")


def main():
    parser = argparse.ArgumentParser(description="Build TensorRT engines from ONNX.")
    parser.add_argument(
        "--onnx-dir",
        type=Path,
        default=Path("models"),
        help="Directory containing ONNX files.",
    )
    parser.add_argument(
        "--output-dir",
        type=Path,
        default=Path("models"),
        help="Directory to store TensorRT engines.",
    )
    parser.add_argument(
        "--fp16",
        action="store_true",
        help="Enable FP16 precision if supported by the device.",
    )
    parser.add_argument(
        "--workspace-gb",
        type=int,
        default=2,
        help="TensorRT workspace size in GB.",
    )
    args = parser.parse_args()

    onnx_files = [
        args.onnx_dir / "superpoint.onnx",
        args.onnx_dir / "lightglue.onnx",
    ]
    for onnx_file in onnx_files:
        if not onnx_file.exists():
            raise FileNotFoundError(f"Missing ONNX file: {onnx_file}")

    for onnx_file in onnx_files:
        engine_file = args.output_dir / (onnx_file.stem + ".plan")
        build_engine(onnx_file, engine_file, fp16=args.fp16, workspace_gb=args.workspace_gb)


if __name__ == "__main__":
    main()