b9848

ggml-org/llama.cppb9848Jun 30, 2026by github-actions[bot]same day as b9847

What changed since b9847

low risk

b9848 adds a CUDA fix for get_rows_back on tables exceeding 65,535 rows; otherwise identical to b9847.

Improvements

  • CUDA get_rows_back now handles tables larger than 65,535 rows via grid-y clamping and adjusted strides.
  • Standard prebuilt binaries refreshed for macOS, Linux, Android, Windows, and openEuler.

Bug fixes

  • Resolved incorrect results/failures in the CUDA backend when gathering rows back from very large tables.

Upgrade notes

This is a single-commit CUDA correctness fix on top of b9847, so no config or API changes are needed. CUDA users working with large embedding or vocabulary tables should upgrade; CPU-only and other backends see no functional difference. KleidiAI remains disabled as in the prior build.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

This release addresses a limitation in the CUDA backend where `get_rows_back` failed for tables exceeding 65,535 rows by implementing grid-y clamping and stride adjustments. Standard binaries are provided for all major platforms.

Key Highlights

  • Fixes `get_rows_back` for tables larger than 65,535 rows in CUDA
  • Includes standard binaries for macOS, Linux, Android, Windows, and openEuler
  • KleidiAI support remains disabled

New Features

  • CUDA large table row handling fix

Release Notes (excerpt)

<details open> CUDA: fix get_rows_back for tables with more than 65535 rows (grid-y clamp + stride) (#25103) </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:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories