b9922

ggml-org/llama.cppb9922Jul 8, 2026by github-actions[bot]same day as b9918

What changed since b9918

low risk

b9922 adds n_keep_tail to split_equal so recurrent models split batches correctly, with no breaking changes.

Improvements

  • split_equal now accepts an n_keep_tail parameter to preserve tail tokens when dividing batches for recurrent models.
  • Batch processing for recurrent architectures is handled more correctly during prompt/slot splitting.
  • Metal f16->f16 GGML_OP_SET_ROWS support from b9918 remains in place; b9922 layers the batch change on top.

Bug fixes

  • Inferred: recurrent models could previously lose or misplace tail tokens when a batch was split, which the new n_keep_tail handling addresses.

Upgrade notes

This is a drop-in rebuild with no config or CLI changes; existing flags and model files work unchanged. Only teams relying on recurrent/hybrid models with large batched prompts will see behavioral differences, and those should be validated against prior outputs.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Adds n_keep_tail support in the split_equal function for better handling of recurrent models during batch processing.

Key Highlights

  • Added n_keep_tail in split_equal for recurrent models
  • Improved handling of recurrent models in batch processing

Release Notes (excerpt)

<details open> llama-batch: add n_keep_tail in split_equal for recurrent models (#25278) </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.2) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:** - Android arm64 (CPU)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories