dx-com 2.3.0

DXCOM Package

DXCOM

DXCOM is a compiler SDK by DeepX that compiles ONNX models into optimized .dxnn binaries for DeepX NPU (Neural Processing Unit).

It performs hardware-aware quantization, graph optimization, and code generation — all in a single pipeline. The compiled .dxnn file can be deployed and executed on DeepX NPU via DX-RT (DeepX Runtime).

Installation

pip install dx-com

Requirements: Python 3.8 – 3.12, Linux (x86_64)

Quick Start

CLI

dxcom -m model.onnx -c config.json -o output/

Argument	Description
-m	Path to the ONNX model file
-c	Path to the JSON configuration file
-o	Output directory for compiled artifacts
--opt_level	Optimization level: 0 (fast) or 1 (higher accuracy, default)
--gen_log	Generate a compilation log
-v	Show version

Python API

import dx_com

dx_com.compile(
    model="model.onnx",
    config="config.json",
    output_dir="output/",
    opt_level=1,
)

Parameters

Parameter	Type	Description
model	str	Path to the ONNX model file
config	str	Path to the JSON configuration file
output_dir	str	Output directory for compiled artifacts
opt_level	int	Optimization level: 0 (fast) or 1 (higher accuracy). Default: 1
calibration_method	str	Calibration method. Options: minmax, percentile, mse, entropy
aggressive_partitioning	bool	Enable aggressive CPU/NPU partitioning. Default: False

Documentation

For detailed technical documentation and user manual, visit DeepX Developer Portal.

License

This software is the proprietary property of DEEPX Ltd. and is provided exclusively to authorized customers with DeepX NPU hardware. See LICENSE for details.