fluttorch 1.0.0

Runtime-agnostic core for shipping PyTorch models to Flutter. Model manifests, tensor specs, drift metrics, and the backend interface.

Changelog

This file records what changed in fluttorch, the contract package. The whole project's record, covering every package and the Python exporter together, is in the repository changelog.

0.7.0

• TensorLayout, so a spec can say which axes are spatial, and DTypeUnsupportedException, which separates a backend with no kernel for an element type from a caller reading bytes as the wrong one. Every exception carries a remedy beside its message.

0.6.0

• TensorSpec can record a layout, nchw or nhwc, so a consumer knows which axes are spatial. Absence is not a default: a spec that says nothing is not a spec that says NCHW, and asking one that does not say raises.
• Every FluttorchException carries a remedy beside its message, and DTypeUnsupportedException separates "this backend has no kernel for that element type" from "the caller read the bytes as the wrong type". The two were one class, and the runtimes reported the first case with the second's message.
• The manifest records which engine executes an artifact, and refuses to be loaded by another one.

0.5.0

• The runtime interface gained the four hooks the ExecuTorch binding implements: a backend pinned at load, repeatable execution, activation taps, and output buffers the caller owns.
• The manifest records the compute precision a delegate was lowered at.

0.4.0

• Quantization recipes and activation specs in the manifest.

0.3.0

• Drift metrics and the tolerance model the parity gate measures against.

0.2.0

• The manifest codec: one document, read identically by Python and Dart.

0.1.0

• Tensor specs, dtypes and the errors this package raises.

0.0.1

• First tag.