Repository navigation
Releases
v1.2.1
Patch release fixing verification accuracy, model conversion edge cases, and safe replacement of existing model bundles.
Fixed
- Verification rejects mismatched output shapes, preserves zero-length input
dimensions and float64 NaN/Inf values, and computes PSNR without overflow or
underflow from squaring errors. Precision warnings now appear in human CLI
output and report the requested PSNR threshold correctly. - Graph cleanup and fusion preserve custom-domain operators. Identity removal
respects shadowed source names in subgraphs, and GELU fusion preserves ranks
introduced by broadcasting constants. - Dynamic
Flatten, scalarGatherND, infinite-inputReduceLogSumExp, and
scalar attention masks now convert and execute correctly. MeanVarianceNormalizationincludes the ONNX-specified1e-9epsilon after
the standard deviation. Some ONNX Runtime versions omit this epsilon; checks
on constant or very small inputs may therefore report a reference discrepancy.- Constant folding checks expanding tensors before allocation. Large constant
RangeandConstantOfShapeoperations remain compact during lowering. - Failed reconversions preserve existing
.aimodelbundles. Replacements are
staged before installation, with rollback and backup retention on I/O errors. - Runtime checks enforce macOS 27+, and
verifyreports malformed model files
asinvalid_model_filerather than a precision-check error. - The MCP extra requires SDK 1.10.0+, which introduced structured tool outputs,
and tool annotations remain compatible with that minimum version.
Validation
1,096 tests passed locally on macOS 27.0.1 with Xcode 27 and coreai-core 1.0.0b3; 19 tests were skipped for documented runtime or reference limitations. Lint, bytecode compilation, wheel/source builds, distribution metadata checks, and package source parity checks passed.
v1.2.0
Version 1.2.0 updates SDK compatibility, fixes operator accuracy, and expands support for dynamic shapes.
Changed
- Support coreai-core 1.0.0b2 and 1.0.0b3, plus Python 3.14.
convert()andOnnxConverter.to_coreai()now return optimized programs. Callsave_asset()directly, or passoptimize=Falseto inspect the IR before optimization.- Keep the MCP SDK dependency below version 2 and report unsupported SDK versions clearly.
Added
- Last-index tie handling for
ArgMaxandArgMin. - Integer
ReduceMeanwith truncation toward zero. - Negative constant padding for cropping, and non-negative padding computed at runtime.
- Dynamic batch and channel dimensions for
Resize, with static spatial dimensions. Concatinputs with different static and dynamic dimension annotations.
Fixed
- Preserve rounding, cast truncation, and integer division through optimization.
- Match ONNX Runtime's float32 resize coordinates and nearest-mode tie handling.
- Avoid scalar gather failures on GPU and handle dynamic-axis
ArgMincorrectly. - Preserve rank-0 tensors when narrowing 64-bit inputs for verification.
- Keep large
Trilumasks compact after optimization. - Resolve the SDK's circular import during converter initialization.
Validation
The full local suite passed on macOS 27 with coreai-core 1.0.0b3: 1,032 tests passed and 17 were skipped for documented limitations. Lint, bytecode compilation, package builds, and distribution metadata checks also passed.
v1.1.2
Changed
- Detect float16 use inside graphs before applying repair rewrites.
- Add provenance metadata when saving bundles.
- Pin NumPy below the next major release pending validation.
Fixed
- Reject invalid split partitions and mismatched lowering result counts.
v1.1.1
v1.1.0
[1.1.0] - 2026-06-15
Changed
- The default
convert/verifyprecision verdict now measures fp32
conversion fidelity. The check still runs first on the runtime's fast
default compute unit, but on a miss it falls back to the deterministic fp32
cpu_onlypath for the verdict (the fallback runs only on a miss, so a
passing conversion pays no extra cost). Two correct-but-divergent classes that
previously failed now pass with a warning: float16 GPU/ANE divergence —
including NaN/overflow and a runtime abort — surfaces as
precision_hardware_divergencewhen the model is faithful on fp32, and benign
large-magnitude accumulation noise (high PSNR) surfaces as
precision_benign_noise.--min-psnrnow sets the benign floor (default
50 dB).result.precision.compute_unitechoes the effective unit behind the
verdict (null= default unit,"cpu_only"= fp32 fallback, or the pinned
unit). An explicit--compute-unitis honored exactly, with no fallback —
pingpu/aneto make the float16 hardware path the verdict.
Added
- Warning codes
precision_benign_noiseandprecision_hardware_divergence
(append-only contract additions; surfaced by the schema dump, CLI, and MCP). verify(..., isolate_execution=...)andOutputReport.elementwise_passed.
Fixed
- The post-conversion precision check now executes the converted
.aimodel
out-of-process. The native Core AI runtime canabort()(not raise) on a
program it cannot execute on the selected compute unit — e.g. an MPSGraph
IOSurfaceassertion on the GPU path — which previously killed the whole CLI
invocation (losing its JSON envelope) or the long-lived MCP server. Such an
abort is now contained and surfaced as a structuredprecision_check_error
(convert exit 3 / verify exit 1); the.aimodelis still written and the
conversion is unaffected. The error message and docs point to
--compute-unit cpu_onlyfor recovery rather than a blind retry.
coreai-onnx 1.0.0
Initial public release.
Highlights:
- ONNX to Core AI .aimodel conversion with 143 built-in lowerings.
- Stable CLI/MCP JSON contract with documented error, warning, and exit codes.
- Validation, precision verification, repair workflow, docs, governance, and PyPI Trusted Publishing configuration.