b10338

ggml-org/llama.cppb10338Aug 10, 2026by github-actions[bot]same day as b10336

What changed since b10336

low risk

Fixes a model-saver key collision that corrupted shared-expert MoE metadata and broke save/load roundtrips.

Improvements

  • Extended test-llama-archs so MoE roundtrip tests use a shared-expert FFN length distinct from n_ff, exercising the previously untested path.
  • Save/load roundtrip coverage now spans all shared-expert MoE architectures rather than a single configuration.

Bug fixes

  • Model saver no longer writes the expert shared FFN length key twice, which had clobbered it with the chunk value and dropped expert_chunk_feed_forward_length entirely.
  • Saved MoE models with shared experts (Qwen2MoE, Qwen3-Next, Granite-MoE, Hunyuan-MoE, ERNIE 4.5, BailingMoE2, Nemotron-H) now reload successfully instead of failing on a shexp tensor-shape mismatch.

Upgrade notes

Upgrade if you save and reload shared-expert MoE models with llama.cpp's built-in saver; GGUF files written by the older build may carry a zeroed shared_feed_forward_length and should be re-converted or re-saved with this version. No config or CLI changes are required, and the WGSL refactor from the prior build carries over unchanged.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Critical bug fix for the model saver that was incorrectly saving MoE (Mixture of Experts) model parameters, causing models like Qwen2MoE to fail loading after a save/load cycle.

Key Highlights

  • Fixed expert shared/chunk FFN length key clobber in model-saver.
  • Resolves loading failures for shared-expert MoE models.
  • Added tests to verify roundtrip save/load for MoE architectures.

Release Notes (excerpt)

<details open> model-saver : fix expert shared/chunk FFN length key clobber (#26693) The saver called add_kv with LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH twice, the second time passing n_ff_chexp. gguf_set_val_u32 removes-then-appends, so the second call clobbers the first: the saved shared_feed_forward_length ends up as n_ff_chexp (0 for every arch except GroveMoE), and expert_chunk_feed_forward_length is never written at all.…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories