b9891

ggml-org/llama.cppb9891Jul 6, 2026by github-actions[bot]same day as b9890

What changed since b9890

low risk

b9891 adds a Metal backend col2im_1d operation for f32/f16/bf16, with no other changes relative to b9890.

New features

  • Metal backend now implements the COL2IM_1D operation for f32, f16, and bf16 tensors, matching CPU/CUDA/Vulkan behavior.

Improvements

  • The Metal supports_op check for COL2IM_1D now requires a contiguous destination and matching source/destination types, aligning it with the other backends.
  • The new kernel uses a single F32 accumulator and one write per output element with no atomics, keeping the implementation simple and deterministic.

Upgrade notes

This is a single-PR increment over b9890, so anyone already on b9890 only gains Metal support for col2im_1d; no config or model changes are needed. The -sm row removal and cuBLAS refactor landed in b9890, so CUDA users upgrading from earlier builds should still account for those changes. Metal users relying on col2im_1d previously fell back to another backend and should now see it run natively.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

A new operation `col2im_1d` has been added to the Metal backend, supporting floating-point formats (f32, f16, bf16), mirroring functionality available in CPU and CUDA backends.

Key Highlights

  • Add col2im_1d operation to Metal backend
  • Support f32, f16, and bf16 formats

Release Notes (excerpt)

<details open> metal: add col2im_1d op (f32/f16/bf16) (#25176) * metal: add col2im_1d op (f32/f16/bf16) Gather kernel mirroring the CPU/CUDA path: each output (t_out, oc) reads its ceil(K/s0) source columns with an F32 accumulator, a single write and no atomics. One thread per output element, 256 per threadgroup. * metal: check dst contiguity and type match in supports_op for COL2IM_1D Align the GGML_OP_COL2IM_1D predicate with the CPU, CUDA, and Vulkan…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories