Skip to content

Add release notes for NNCF v3.1.0 #129

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
Copilot wants to merge 3 commits into
base: seo_skils
Choose a base branch
from
Draft

Add release notes for NNCF v3.1.0 #129

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c91d5ef
Initial plan
Copilot Apr 1, 2026
dca6b5c
Initial plan for release notes generation
Copilot Apr 1, 2026
4a3e2cf
Add release notes for v3.1.0
Copilot Apr 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,3 +138,4 @@ nncf-tests.xml
compressed_graph.dot
original_graph.dot
tests/post_training/**/*memory_logs
tmp_release_commits.txt
35 changes: 35 additions & 0 deletions ReleaseNotes.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,40 @@
# Release Notes

## New in Release 3.1.0

Post-training Quantization:

- Features:
- (OpenVINO) Introduced NVFP4 (f4e2m1) data type for weight compression. Uses f4e2m1 precision with constant group size 16, and compresses the scale to f8e4m3 with a second-degree FP32 scale. (#3967)
- (OpenVINO) Added `backup_mode` parameter for FP compression formats (FP4, FP8, MXFP4, MXFP8), allowing first and last layers to use an FP backup format instead of INT8. (#3886)
- (PyTorch) Added experimental converter for NNCF-compressed linear modules to GPTQModel format with Triton kernel support. (#3848)
- (PyTorch) Added `TopKMetatype` support to fix graph building for models with TopK operations, such as YOLO26. (#3944)
- Added lazy import of `nncf.torch` in `nncf/__init__.py`, allowing `nncf.torch.*` API to be accessible without explicit `import nncf.torch`. (#3862)
- Fixes:
- (OpenVINO) Fixed incorrect usage of `do_float_quantization` after NVFP4 introduction that caused wrong compressed weights in the scale estimation algorithm. (#3991)
- (OpenVINO) Fixed scale estimation for the adaptive codebook compression type for improved accuracy. (#3888)
- (OpenVINO) Extended RoPe ignored pattern to cover cases without a transpose node, fixing weight compression for the Phi-3.5-moe model. (#3989)
- (ONNX) Fixed `nncf.errors.ValidationError` about missing tensor names during ONNX model transformations. (#3988)
- Improvements:
- Migrated internal `NNCFGraph` from `nx.DiGraph` to `nx.MultiDiGraph` to correctly represent and quantize models with parallel edges (e.g., YOLO26 SDPA blocks, RoPE patterns with parallel inputs). (#3843)
- Added documentation section explaining the gradient computation formula for Quantization-Aware Training. (#3965)
- Added Olive to the NNCF integrations section in documentation. (#3998)

Deprecations/Removals:

- Removed outdated scripts from the `tools` directory. (#3939)

Requirements:

- Updated OpenVINO to 2026.1.0rc2. (#4005)
- Updated PyTorch (2.10.0), TorchAO (0.16.0), and Torchvision (0.25.0) versions. (#3852)
- (PyTorch) Migrated from `torch.ao` to the standalone `torchao` package. (#3854)
- (ONNX) Updated onnxruntime from 1.21.1 to 1.24.3. (#3977)
- (ONNX) Updated onnx from 1.17.0 to 1.20.1. (#3966)
- Updated numpy upper bound to <2.5.0. (#3875)
- Moved pandas to optional dependencies. (#3970)
- Removed unused pillow dependency. (#3929)

## New in Release 3.0.0

Post-training Quantization:
Expand Down