b10327

ggml-org/llama.cppb10327Aug 8, 2026by github-actions[bot]same day as b10326

What changed since b10326

low risk

b10327 fixes CUDA quantized-copy kernel launch geometry; no API or behavior changes beyond that.

Improvements

  • Quantized copy kernels on CUDA now launch with correctly computed thread and block counts, avoiding malformed grid configurations.
  • Test coverage extended with an uneven block count copy case to guard against regressions in kernel launch sizing.

Bug fixes

  • Corrected the thread/block count calculation in CUDA quantized copy kernel launches, which could previously mis-size the grid.

Upgrade notes

This is a single targeted CUDA fix on top of b10326, with no config, CLI, or model-format changes. CUDA users doing quantized copies (e.g. KV-cache or tensor copies) should upgrade; CPU-only and other backend users gain nothing but face no risk. No migration steps required.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Fixes a bug in CUDA kernel launches for quantized copy operations where the thread/block count was incorrectly calculated.

Key Highlights

  • Corrected thread/block count logic in quantized copy kernels to ensure proper kernel execution.
  • Added tests for uneven block count scenarios to validate the fix.

Release Notes (excerpt)

<details open> CUDA: fix thread/block count in quantized cpy kernel launches (#26731) * CUDA: fix thread/block count in quantized cpy kernel launches * tests: add uneven block count cpy case </details> **Website:** - <https://llama.app> **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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories