b10235

ggml-org/llama.cppb10235Aug 2, 2026by github-actions[bot]same day as b10234

What changed since b10234

low risk

Metal backend gains an F32 SILU_BACK kernel, a small incremental step beyond b10234's F16 binary-op support.

New features

  • Metal backend now implements the SILU_BACK operation for F32 precision.

Improvements

  • Removed redundant assertions in the Metal SILU_BACK op path, simplifying the kernel dispatch code.
  • Numerical correctness and stability improved for training-style backward passes on Apple GPUs.

Upgrade notes

This is a build-number bump with a single Metal kernel addition; no config, model, or CLI changes are involved. Apple Silicon users running fine-tuning or backward-pass workloads benefit, while other backends are unaffected. Rebuild or pull the new prebuilt binaries as usual.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Implemented the SILU_BACK (Sigmoid Linear Unit Backward) operation for the Metal backend to support F32 precision.

Key Highlights

  • Implemented SILU_BACK operation for Metal F32.
  • Removed redundant asserts in the `ggml_metal_op_silu_back` function.
  • Improves numerical stability and correctness.

Release Notes (excerpt)

<details open> metal : add SILU_BACK (#25982) * feat(silu_back): implemented silu_back op for f32 * fix(silu_back): removed redundant asserts in ggml-metal-ops.cpp function ggml_metal_op_silu_back. </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