b10549

ggml-org/llama.cppb10549Aug 21, 2026by github-actions[bot]same day as b10545

What changed since b10545

medium risk

Adds tensor-parallel tensor-split support for LFM2 and LFM2MOE models; no other changes noted between builds.

New features

  • Tensor split is now supported for LFM2 and LFM2MOE architectures when running in Tensor Parallelism mode.

Improvements

  • Multi-GPU TP deployments can now shard LFM2-family models instead of falling back to single-device execution.

Upgrade notes

No configuration or migration steps are required; the change only affects users running LFM2/LFM2MOE with tensor parallelism enabled. If you previously avoided TP for these models due to unsupported splitting, you can now enable it and validate output quality against a single-device baseline. The Metal K-extent fix from the prior build is unchanged and carries over.

Analyzed Oct 26, 2026

Upgrading?

Release notes summary

Enables tensor split support for LFM2 and LFM2MOE models in Tensor Parallelism (TP) mode.

Key Highlights

  • TP: enable tensor split for LFM2/LFM2MOE models
  • Assisted by deepseek-v4-flash

Release Notes (excerpt)

<details open> TP: enable tensor split for LFM2/LFM2MOE (#26993) Assisted-by: deepseek-v4-flash </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/42096995> **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) - Ubuntu arm64 (Vulkan)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories