b9661

ggml-org/llama.cppb9661Jun 16, 2026by github-actions[bot]same day as b9660

What changed since b9660

low risk

b9661 adds the col2im_1d op to the Vulkan backend, following the existing CPU implementation.

New features

  • Vulkan backend now supports GGML_OP_COL2IM_1D, matching the previously added CPU operation.

Improvements

  • col2im_1d Vulkan kernel uses a bounded gather loop instead of a full-K scan with modulo, reducing wasted work.
  • Unsupported types now return nullptr in the Vulkan col2im_1d path rather than attempting an invalid dispatch.

Upgrade notes

No config or API changes are required; this is a drop-in binary swap. Vulkan users gain acceleration for models that use col2im_1d, while CPU and other backends are unaffected. The LFM2 tool-call escaping fix from b9660 remains in place.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Implements the col2im_1d operation in the Vulkan backend, following up on the CPU implementation. This adds a new mathematical operation to the Vulkan acceleration layer.

Key Highlights

  • Add col2im_1d operation to Vulkan backend
  • Improved gather loop logic
  • Standard release builds for all platforms
  • Disabled KleidiAI build for macOS Apple Silicon

Release Notes (excerpt)

<details open> vulkan: add col2im_1d op (#24425) * vulkan: add GGML_OP_COL2IM_1D, follow-up to the CPU op * vulkan: col2im_1d bounded gather loop instead of full-K scan with modulo * vulkan: col2im_1d address review from @jeffbolznv * vulkan: col2im_1d return nullptr for unsupported types, address review from @0cc4m </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories