b10152

ggml-org/llama.cppb10152Jul 27, 2026by github-actions[bot]same day as b10151

What changed since b10151

low risk

b10152 fixes GPU layer offload accounting so MTP (nextn) blocks no longer push front layers off the GPU.

New features

  • n_gpu_layers now accounts for nextn/MTP blocks when deciding how many layers fit on the GPU.

Improvements

  • Layer-fitting logic keeps the model's front layers resident on the GPU for MTP-enabled models instead of misallocating them.
  • Build-time SYCL ocloc parallelization from the previous build remains in place, so no build regression is expected.

Bug fixes

  • Corrected offload layer counting that previously caused front layers to be dropped to CPU when MTP blocks were present.

Upgrade notes

Drop-in replacement for b10151; no config or CLI changes are required. Users running MTP/nextn models with an explicit -ngl value should re-check GPU memory headroom, since more layers may now be placed on the GPU than before.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Optimizes GPU layer allocation for Multi-Threaded Parallel (MTP) blocks to ensure the front layers of the model remain on the GPU when using the n_gpu_layers parameter.

Key Highlights

  • Optimized MTP block counting for GPU layer management
  • Available binaries for macOS, Linux, Windows, Android, and openEuler

Release Notes (excerpt)

<details open> fit : count nextn (MTP) blocks in n_gpu_layers so front layers stay on GPU (#26177) </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) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.2) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories