b9736

ggml-org/llama.cppb9736Jun 20, 2026by github-actions[bot]same day as b9735

What changed since b9735

low risk

b9736 fixes GLM-5.2 GGUF loading by making DSA indexer tensors optional, on top of b9735's Intel AMX speedups.

New features

  • GLM-5.2 GGUFs with the DSA lightning indexer present on only a subset of layers can now be loaded successfully.

Improvements

  • Indexer tensors are now marked as not-required in the GLM_DSA loader, so layers lacking them resolve to nullptr instead of failing.
  • Retains the prior release's AMX quantization optimization for Intel Xeon, which flattened partitioning across n_batch * M for reported speedups up to ~1.47x.

Bug fixes

  • Resolved 'missing tensor' load failures (e.g. blk.3.indexer.k_norm.weight) when loading partial-indexer GLM-5.2 models.
  • Models without indexer layers now fall back cleanly to full MLA attention rather than aborting at load time.

Upgrade notes

No configuration or migration steps are required; this is a drop-in binary swap. Anyone running GLM-5.2 GGUFs should upgrade, since earlier builds refused to load them. DeepSeek-V3.2 behavior is unchanged, and the indexer tensors remain loaded-but-unused because indexer runtime is not yet implemented.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Fixes loading of GLM-5.2 GGUF models by marking DSA indexer tensors as optional, allowing models with partial indexer layers to load correctly as full MLA attention.

Key Highlights

  • GLM-5.2 DSA loader now supports optional indexer tensors
  • Fixes missing tensor errors for partial indexer layers
  • DeepSeek-V3.2 remains unaffected

Release Notes (excerpt)

<details open> model : glm-dsa load DSA indexer tensors as optional (#24770) GLM-5.2 ships the DSA "lightning indexer" on only a subset of layers (the "full" layers; others omit it), but the GLM_DSA loader created the five indexer tensors on every layer as required, so loading any GLM-5.2 GGUF failed with e.g. `missing tensor 'blk.3.indexer.k_norm.weight'`. GLM_DSA's graph is llama_model_deepseek2::graph (plain MLA) and does not use…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories