b9689

ggml-org/llama.cppb9689Jun 17, 2026by github-actions[bot]same day as b9688

What changed since b9688

low risk

Metal backend concat operator gains f16, bf16, i8, i16 and i64 tensor type support.

New features

  • Metal concat operator now accepts f16 and bf16 tensors in addition to the previous f32 and i32 support.
  • Concat on Metal extended further to i8, i16 and i64 integer tensor types.

Improvements

  • Concat kernel is now templated on element type with specializations for float, half, bfloat and int, replacing the fixed-type implementation.
  • Pipeline lookup refactored into a type-aware getter so the correct concat specialization is selected per tensor type.
  • Device capability check now permits f16 unconditionally and gates bf16 on hardware bfloat16 support.

Upgrade notes

This is a Metal-only backend capability expansion with no API or config changes; existing models and flags continue to work unchanged. Apple Silicon users running graphs that previously fell back to CPU for non-f32/i32 concat may see those ops stay on GPU. No migration steps are required.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Extends the Metal backend concat operator to support a wider range of tensor types including f16, bf16, i8, i16, and i64.

Key Highlights

  • Metal concat operator support for f16 and bf16
  • Template kernel specialization
  • Device support checks for bfloat16

Release Notes (excerpt)

<details open> metal : add f16 and bf16 support for concat operator (#24724) * metal : add f16 and bf16 support for concat operator Extend the Metal backend concat operator to support f16 and bf16 tensor types in addition to the existing f32 and i32 support. - Template kernel_concat on type T with specializations for float, half, bfloat, and int - Add type-specific pipeline getter ggml_metal_library_get_pipeline_concat()…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories