pypi / ai-edge-quantizer

ai-edge-quantizer

What changed in ai-edge-quantizer 0.4 to 0.10, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.

3 breaking11 new APIs and options5 performance2 behavior and defaults1 tooling6 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

ai-edge-quantizer 0.103 changes

Adds an experimental algorithm for FULLY_CONNECTED that combines activation-aware per-channel activation scaling with optimal weight clipping.
0.10.0release notes
Adds uint8 data type support for model INPUT/OUTPUT tensors in recipes, enabling compatibility with some hardware accelerators.
0.10.0release notes

Recipe validation no longer requires weight_tensor_config for ops without weights, such as model INPUT/OUTPUT, so dummy weight configs are unnecessary.
0.10.0release notes

Same-size Hadamard matrices now use a single tensor, and inserted Hadamard matrices are float.
0.9.0release notes

Recipes and buffer packing were added for 2-bit model quantization.
0.9.0release notes
Weight recovery algorithms now support scale recovery and blockwise quantization, with ready-to-use 4-bit and 8-bit named recipes.
0.9.0release notes
Calibrator and Quantizer gained a profiler-based calibration mode for faster, more memory-efficient XNNPACK-assisted calibration.
0.9.0release notes
A calibration API was added that lets developers reuse the inference pipeline for calibration.
0.9.0release notes
GPTQ support was added as a quantization algorithm.
0.9.0release notes
Int4 matrix multiplication Hadamard transforms gained a max_hadamard_size control.
0.9.0release notes
LiteRT-LM files can now be read, quantized, and written, including extracting serialized TFLite models and applying recipes across TFLiteModel sections.
0.9.0release notes
Model validation now supports multiple metrics.
0.9.0release notes

Large tensors are processed in chunks along the leading dimension to avoid generating large intermediate values.
0.9.0release notes
These buffers are explicitly garbage-collected immediately after output buffer writing to reclaim memory.
0.9.0release notes
Model serialization uses disk-backed memory mappings and chunk-by-chunk file writing to prevent OOM errors on large LLMs.
0.9.0release notes
Buffer mapping IDs are tracked, producer and consumer ops are pre-computed, and empty buffers are pruned to reduce serialized file size and parsing overhead.
0.9.0release notes
Quantization uses zero-copy array views and in-place core operations instead of unnecessary copies, lowering peak memory usage.
0.9.0release notes

A CLI was introduced for running model quantization workflows from the terminal.
0.9.0release notes

Calibration metadata tracking was added.
0.9.0release notes
Crashes in float casting operations with non-constant weights were fixed.
0.9.0release notes
Output tensors are no longer inadvertently rotated during fully connected op processing.
0.9.0release notes
Accidental double-quantization of input tensors is prevented.
0.9.0release notes
Progress bars and reports were added for quantization steps and calibration passes.
0.9.0release notes

Bias error checks are disabled.
0.4.2release notes
Python 3.9 support was dropped and Python 3.13 support was added.
0.4.2release notes