b10361
ggml-org/llama.cppb10361Aug 11, 2026by github-actions[bot]same day as b10360
What changed since b10360
b10361 fixes SWA being silently disabled for EXAONE 4.5 models and adjusts TENSOR_SKIP handling in the metadata-only loader path.
New features
- EXAONE 4.5 GGUFs carrying an MTP/nextn head now load with Sliding Window Attention actually enabled instead of falling back to no SWA.
- Added test coverage for exaone4 hparams ordering using a synthetic model laid out like the shipped EXAONE 4.5 GGUFs.
Improvements
- Model loader no longer asserts on a null buffer type when constructing TENSOR_SKIP-marked tensors from metadata alone, matching the file-backed path's behavior.
- Hparams reading order corrected so n_layer_nextn is populated before architecture checks that depend on n_layer().
Bug fixes
- Fixed SWA not being enabled for EXAONE 4.5 because the n_layer()==64 check ran before the nextn layer count was read, making block_count 65 evaluate as 65.
- Fixed llama_model_init_from_user failing outright on architectures that mark nextn/MTP layers with TENSOR_SKIP.
- Reverted the earlier TENSOR_SKIP-related changes that this fix supersedes.
Upgrade notes
No configuration or CLI changes are required; this is a drop-in replacement for b10360. Anyone serving EXAONE 4.5 (especially the official LGAI-EXAONE GGUFs) should upgrade, since prior builds reported a plausible n_swa value in logs while SWA was effectively off, silently degrading long-context behavior. Re-verify output quality on EXAONE 4.5 after upgrading, as attention behavior changes.
Analyzed Oct 28, 2026
Upgrading?
Release notes summary
Fixed a critical bug where Sliding Window Attention (SWA) was not being enabled for EXAONE 4.5 models due to incorrect parameter reading order.
Key Highlights
- Fixed SWA not being enabled for EXAONE 4.5
- Resolved hparams ordering bug (n_layer vs n_layer_nextn)
- Reverted previous TENSOR_SKIP related changes
Release Notes (excerpt)
<details open> model : fix SWA not being enabled for EXAONE 4.5 (#26848) * model : fix SWA not being enabled for EXAONE 4.5 load_arch_hparams tests `hparams.n_layer() == 64` before LLM_KV_NEXTN_PREDICT_LAYERS has been read. n_layer() returns n_layer_all - n_layer_nextn and n_layer_nextn defaults to 0, so a GGUF carrying the MTP head (block_count=65, nextn=1) evaluates to 65 and the whole SWA block is skipped. The model type switch further down in the…
Published by ggml-org on GitHub.
Read the full release notesRelated Repositories
:electron: Build cross-platform desktop apps with JavaScript, HTML, and CSS
Godot Engine – Multi-platform 2D and 3D game engine
ClickHouse® is a real-time analytics database management system
Self-hosted game stream host for Moonlight.
DuckDB is a high-performance analytical database system