pypi / nncf

nncf

What changed in nncf 3, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.

13 breaking6 new APIs and options3 behavior and defaults10 fixes and other

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

npx @unstale/cli init

nncf 332 changes

Removed the unused PyTorch CompressionParameter class.
3.3.0release notes
Removed the unused NNCFDataLoader class.
3.3.0release notes
Removed the unused PyTorch patch_torch_operators function.
3.3.0release notes
Removed the unused PyTorch PTInitializingDataLoader.
3.3.0release notes
Removed the unused QuantizationStatistics class.
3.3.0release notes
Removed the unused QuantizersCounter class.
3.3.0release notes
Removed the unused PyTorch register_module function.
3.3.0release notes
Removed the unused Statistics class.
3.3.0release notes
The deprecated INT8 mode has been removed; use INT8_SYM or INT8_ASYM instead.
3.2.0release notes
PyTorch PTQ and Weight Compression algorithms remove no-op operations such as to and dropout from the inference graph.
3.2.0release notes
The unused example_input parameter has been removed.
3.2.0release notes
OVQuantizer has been removed from NNCF; use it from the executorch package instead.
3.2.0release notes
The unused ProgressBar class has been removed.
3.2.0release notes

AWQ and Scale Estimation support GroupedMatMul, enabling data-aware weight compression of grouped_mm-based MoE models.
3.4.0release notes
Added OpenVINO GroupedMatMul operation support for data-free weight compression.
3.3.0release notes
Added INT2 symmetric compression type support for OpenVINO.
3.3.0release notes
Added INT3 symmetric compression type support for OpenVINO.
3.3.0release notes
Can now ignore specific weight nodes by constant node name in nncf.compress_weights and nncf.prune.
3.2.0release notes
The Scale Estimation algorithm supports the transpose_a attribute in ONNX and OpenVINO backends.
3.2.0release notes

The ignored pattern recognizes the matmul → transpose → cos/sin structure used by models such as GPT-OSS.
3.4.0release notes
AWQ, GPTQ, Scale Estimation, and LoRA Correction are enabled for all bit widths below 4, including INT2 and INT3.
3.3.0release notes
Added FP8 support for the ONNX backend.
3.3.0release notes

Fixes a KeyError when resolving a bias attribute for a tensor that is a model input and has no parent node.
3.4.0release notes
The node following split-like operations such as chunk is removed from the TorchFX inference graph, fixing statistic collection errors for models such as YOLO11.
3.4.0release notes
Fixes squeeze axis determination for ONNXEmbeddingMetatype nodes.
3.4.0release notes
Fixed the compression range when all weight values are positive or all are negative.
3.3.0release notes
Fixed the aggregator formula used in mixed-precision weight compression.
3.3.0release notes
Weight serialization supports safetensors >= 0.8.0 by converting NumPy scalars to arrays before saving.
3.3.0release notes
Scale Estimation now properly handles zero initial_steps and scale_steps.
3.2.0release notes
Serialized config now saves the module path, improving load_from_config reliability.
3.2.0release notes
Fixed unexpected behavior when loading quantization constraints from a config dictionary.
3.2.0release notes
Prevents capture of nodes that do not affect target outputs.
3.2.0release notes