b10174
ggml-org/llama.cppb10174Jul 29, 2026by github-actions[bot]same day as b10173
What changed since b10173
b10174 adds GLM-5.2 NextN/MTP speculative decoding plus --mtp/--no-mtp conversion export options, on top of b10173's Laguna-S-2.1 model type.
New features
- GLM-5.2 (GLM_DSA) can now be used as a --spec-type draft-mtp target for NextN/MTP speculative decoding.
- The conversion tool exposes --mtp/--no-mtp flags for GlmMoeDsaForCausalLM, letting users bundle or strip the NextN block (blk.78) at export time.
- MTP draft head is built via a dedicated graph_mtp path (enorm/hnorm/eh_proj, dense MLA, sigmoid-gated MoE with shared expert and shared head).
Improvements
- MTP context now uses a plain attention KV cache limited to the nextn layers, while the main context keeps the DSA cache filtered to trunk layers only.
- NextN tensors are loaded using the qwen35moe/step35-style presence probe, reusing existing loading conventions.
- MTP graph passes _s scale tensors so NVFP4-quantized GLM-5.2 draft heads work correctly.
Upgrade notes
No configuration or CLI changes are required for existing workloads, and default conversion output for GLM-5.2 is unchanged, so previously converted models remain valid. To enable speculative decoding, re-export the model with --mtp and launch with --spec-type draft-mtp; use --no-mtp to keep the prior trunk-only export. Laguna-S-2.1 support from b10173 is unaffected.
Analyzed Nov 1, 2026
Upgrading?
Release notes summary
Added speculative decoding support (NextN/MTP) for GLM-5.2 (GLM_DSA) and updated the conversion tool to support `--mtp/--no-mtp` export options for this model.
Key Highlights
- Added NextN/MTP speculative decoding support for GLM-5.2
- Added conversion support for --mtp/--no-mtp export for GLM-5.2
Release Notes (excerpt)
<details open> model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2) (#25980) * model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2) Adds GLM-5.2 NextN/MTP as a --spec-type draft-mtp target: nextn tensor loading via the qwen35moe/step35-style presence probe, a graph_mtp builder (enorm/hnorm/eh_proj + dense MLA + sigmoid-gated MoE with shared expert + shared head with fallbacks, _s scale tensors passed…
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