v0.5.12

pywebnn v0.5.12

Changes

• Pin rustnn to "0.5.12".
• Pin webnn-graph to "0.3".
• Update to new ONNX split protobuf + weight API and MLOperandDataType enum (#14).
• Remove the use of shape_inference OperatorOptions (#13).
• ci: install protoc in the publish workflow so webnn-onnx-utils proto compilation works on the macOS runner.

Validation

Built and verified locally against the released rustnn 0.5.12 + webnn-graph 0.3.0 + webnn-onnx-utils 0.1.1 from crates.io. All 5 e2e demos passed:

1. Quantization round-trip
2. MiniLM embeddings (Hugging Face Hub)
3. MobileNetV2 image classification (Hugging Face Hub)
4. SmolLM-135M text generation (Hugging Face Hub)
5. KV cache with device tensors

Notes

• Wheels are built and published to PyPI by the Publish to PyPI workflow on this release.
• The WebNN frontend (web feature) is intentionally not exposed in this release; it will return in a separate rustnn-web workspace crate (tracked in rustnn).