●Change8

b10174

📦 llama-cppView on GitHub →
✨ 6 features🔧 2 symbols

Summary

This release introduces NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2) and enables --mtp/--no-mtp export for GlmMoeDsaForCausalLM. It also includes various pre-compiled binaries for different platforms and hardware accelerators.

Migration Steps

  1. For GLM-5.2 models, when exporting with --no-mtp, the appended NextN block (blk.78) and its nextn_predict_layers KV will be dropped.
  2. For GLM-5.2 models, when exporting with --mtp, only the NextN block plus shared embeddings/norm/lm_head will be kept.

✨ New Features

  • Added NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2).
  • Implemented NextN tensor loading via qwen35moe/step35-style presence probe.
  • Introduced a graph_mtp builder for GLM_DSA.
  • Added t_h_nextn extraction in the trunk graph.
  • Implemented MTP-context KV setup for the draft head.
  • Enabled --mtp/--no-mtp export for GlmMoeDsaForCausalLM (GLM-5.2).

Affected Symbols

GLM_DSAGlmMoeDsaForCausalLM