b10232

ggml-org/llama.cppb10232Aug 2, 2026by github-actions[bot]same day as b10231

What changed since b10231

low risk

b10232 adds Metal GPU kernels for DeepSeek V4 hyper-connections, on top of b10231's DSpark sidecar auto-resolution.

New features

  • Metal backend now implements the DSV4 hyper-connection ops (COMB, PRE, POST), enabling DeepSeek V4 hyper-connection layers to run on Apple GPUs.
  • New GGML operations GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE and GGML_OP_DSV4_HC_POST are exposed with Metal dispatch plumbing.
  • DSpark speculative sidecars (carried over from b10231) are auto-discovered and selected, outranking dflash when no type is specified.

Improvements

  • Hyper-connection kernels use SIMDgroup registers and shuffle operations for better Metal throughput.
  • Sinkhorn iteration count and embedding width are now validated against production values in tests.

Upgrade notes

This is a single-commit build bump with purely additive Metal backend support; no CLI flags, model formats or config files change. macOS/Apple Silicon users running DeepSeek V4 hyper-connection models get GPU acceleration, while other backends and platforms are unaffected. Rebuild or pull the new binaries as usual; nothing to migrate.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Implemented optimized Metal kernels for DeepSeek V4 hyper-connections using SIMDgroup registers and shuffle operations.

Key Highlights

  • Implemented GGML_OP_DSV4_HC_COMB, PRE, and POST operations.
  • Optimized kernels using SIMDgroup registers and shuffle.
  • Added Metal dispatch plumbing and tested Sinkhorn iteration count.

Release Notes (excerpt)

<details open> metal: implement DeepSeek V4 hyper-connections (#26459) - Implement GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and GGML_OP_DSV4_HC_POST with SIMDgroup register and shuffle optimized kernels. - Add Metal dispatch and support plumbing and test the production Sinkhorn iteration count and embedding width. Assisted-by: Codex Co-authored-by: Thiago Padilha <thiago@padilha.cc> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories