onnxnpu 0.1.1

CLI toolkit to check, optimize and modify ONNX models for NPU deployment

ONNXNPU Toolkit

onnxnpu is a lightweight CLI toolkit that inspects, optimizes, and modifies ONNX models before you convert or deploy them to NPUs.

"Catch unsupported operators early before they derail your model." — Mason Huang

✨ Features (v0.1 — released)

Feature	Description
Operator scan	Fast, dependency‑free static analysis of .onnx files
Hardware profiles	Built‑in JSON compatibility tables for common NPUs (KL520 / 530 / 630 / 720 / 730 …) with an override mechanism
Clear report	CLI table ＋ optional JSON / Markdown export; highlights unsupported ops and optional‑feature gaps
Actionable hints	Suggestions and links to official docs for each unsupported operator
Opset update	Upgrade model opset (12 – 18) to match target hardware

---

🧭 Roadmap

Version	Target Date*	Major Items	Notes / Dependencies
0.2 – Validation & Reporting	May 2025	• Shape checker enforcing 4‑D (1, C, H, W) constraint • Rich Markdown / JSON report templates for CI badges	Uses ONNX shape‑inference to avoid manual parsing
0.3 – Graph Simplification & Slimming	Jun 2025	• Integrate onnx‑sim (--simplify) • Model slimming (--prune, --quantize) • Bundle Kneron optimizer_scripts (BN‑Conv fuse, Dropout removal …)	Requires onnx‑sim ≥ 0.4; quantization via ONNX QOps
0.4 – Automatic Op Replacement	Jul 2025	• --replace mapping table (e.g., Reshape → Flatten) • Fallback to custom kernels / plugin stubs	Needs rule set ＋ regression tests
0.5 – Interactive Viewer	Aug 2025	• onnxnpu view drag‑and‑drop web UI • Highlight unsupported nodes directly on the graph • Downloadable HTML report	Likely React + ONNX‑JS; demo hosted on GitHub Pages
0.6 – Extensibility & Ecosystem	Sep 2025	• Plugin system via Python entry‑points • Community hardware‑profile submission flow • Freeze stable API v1.0	Plan to publish on conda‑forge after API stabilisation

* Dates are tentative and may shift based on resources.

🚀 Quick start

You can use two different CLI commands: onnxnpu or onpu to check, optimize, and modify your ONNX models for NPU deployment. Both commands provide identical functionality with the same syntax.

Command	Description
pip install onnxnpu	Install latest package from PyPI
onnxnpu check my_model.onnx -p KL720	Check my_model.onnx for the KL720 hardware profile
onnxnpu check my_model.onnx	Check my_model.onnx for all built-in profiles
onnxnpu opt my_model.onnx --opset 13	Update model to opset 13
onnxnpu -V, onnxnpu --version	Show version number and exit

Sample output

Model summary – my_model.onnx
IR version : 6    Opset : 11
Inputs  : input  float32  [1, 3, 112, 112]
Outputs : output  float32  [1, 512]
Dynamic axes : None detected ✓

my_model.onnx · IR 6 · KL520
╭──────────────────────────────────────────────────────────────╮
│  Status  Operator   Count   Notes                            │
├──────────────────────────────────────────────────────────────┤
│   ✓      Conv        27                                      │
│   ✓      Relu        27                                      │
│   ✗      Elu          5     Not supported on KL520           │
│   ✓      MaxPool      5                                      │
│   ✗      Resize       2     Only linear/nearest modes OK     │
╰──────────────────────────────────────────────────────────────╯
⚠  2 unsupported operator(s) detected.

🧑‍💻 API usage

from onnxnpu import Checker, load_profile

checker = Checker("my_model.onnx", profile=load_profile("kl720"))
report = checker.run()
print(report.to_markdown())

# Update opset version
from onnxnpu import update_opset_version
update_opset_version("my_model.onnx", target_version=13)

📖 Hardware profiles

Profiles live under onnxnpu/profiles/*.json. Each profile declares the operators, attributes, and constraints supported by a particular accelerator. See docs/PROFILE_SCHEMA.md for the JSON schema.

Contributions for new hardware are very welcome!

A note on language

The primary language of this README is English for wider community reach. A Traditional Chinese translation will be added soon.