b10247

ggml-org/llama.cppb10247Aug 4, 2026by github-actions[bot]same day as b10246

What changed since b10246

low risk

b10247 replaces fixed-size split-graph input arrays with dynamic allocation, fixing crashes on wide MoE models.

New features

  • Backend scheduler now grows split input and graph input buffers on demand instead of using fixed-size arrays.
  • Graph size is computed from the actual input tensor count rather than a hardcoded constant.

Improvements

  • Removes the 30-input ceiling that previously limited graph splits in multi-backend scheduling.
  • Enables loading of wide MoE architectures (Gemma 4, Qwen MoE, Mixtral, DeepSeek) that previously exceeded the fixed buffer.
  • Internal cleanup of the scheduler allocation code following the refactor.

Bug fixes

  • Fixes crashes when loading wide MoE models on multi-backend setups where graph splits exceed 30 input tensors.

Upgrade notes

This is a drop-in build bump with no config or API changes required. Users running large MoE models across multiple backends should upgrade to avoid the split-input crash; the prior OpenCL q6_K lm_head optimization from b10246 remains in place.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Replaced fixed-size arrays with dynamic allocation for split graph inputs to fix crashes when loading wide MoE models (Gemma 4, Qwen MoE, Mixtral, DeepSeek).

Key Highlights

  • Replaced fixed-size arrays with dynamic allocation for split graph inputs
  • Fixes crashes when loading wide MoE models on multi-backend setups
  • Graph size calculation now uses actual input count instead of fixed constant

Release Notes (excerpt)

<details open> ggml: use dynamic allocation for split graph inputs (#22789) * ggml: use dynamic allocation for split graph inputs Replace fixed-size GGML_SCHED_MAX_SPLIT_INPUTS arrays with dynamically allocated buffers in the backend scheduler. This fixes crashes when loading wide MoE models (Gemma 4, Qwen MoE, Mixtral, DeepSeek) on multi-backend setups where graph splits exceed 30 input tensors. - split->inputs: dynamic array with grow-on-demand…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories