[Mirror] New quantization type: Q3_HIFI#65
Open
ngxson wants to merge 295 commits intongxson:masterfrom
Open
Conversation
All 3 metrics beat Q3_K_M
Q3_HIFI_A now the official version
Speed benchmark script added
Latest updates
Latest updates
Q3_HIFI model ready to go
Q2_K_HIFI updates
Q5_K_HIFI speed improvements
… add whitespace for clarity in CPU quantization functions
LITE variant
…fy output weight matching and streamline default type returns for various quantization formats.
- Refactor `.editorconfig` to apply settings for public tools and remove deprecated entries. - Add `.gitattributes` to treat the WebUI build as binary for diff purposes. - Update `.gitignore` to exclude gzipped index.html files. - Introduce `AGENTS.md` and `CONTRIBUTING.md` to clarify AI usage policies and contribution guidelines. - Modify `CMakeLists.txt` to include an option for building the embedded Web UI. - Update `README.md` with recent API changes and hot topics. - Adjust Dockerfiles and other configuration files for compatibility and dependency updates. - Add new model support in `convert_hf_to_gguf_update.py` and related scripts. These changes enhance project maintainability and clarify contribution expectations.
- Introduce fallback defaults for layer normalization and rope frequency base if not specified in hparams. - Remove unnecessary addition of BOS token in Gemma4Model. - Ensure compatibility with updated model parameter requirements. These changes improve the robustness of model parameter handling and align with the latest specifications.
The previous refactoring removed explicit parameter writes from Gemma3Model.set_gguf_parameters() and Gemma4Model.set_vocab(), assuming TextModel would handle them. This broke Gemma4 inference because the GGUF file's key ordering matters: gguf_find_key() returns the *first* occurrence of a key, so TextModel writing head_count_kv as a scalar first caused all 60 layers to use n_head_kv=1 instead of the correct per-layer array (1 for SWA, 4 for global). This produced all-zero logits and PPL == n_vocab == 262144 for every imatrix chunk. Restore Gemma3Model to explicitly write all parameters (matching upstream exactly), and restore add_add_bos_token(True) in Gemma4Model. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Mirror from upstream PR: ggml-org#18246
Summary by CodeRabbit
New Features
Runtime / Backend
Documentation
Testing & Tools
Chores