b10150

ggml-org/llama.cppb10150Jul 27, 2026by github-actions[bot]same day as b10149

What changed since b10149

medium risk

b10150 reworks how ops are offloaded to a weight's backend and fixes DeepSeek-V4 graph execution.

New features

  • Operation placement is now decided per-weight based on which backend holds that weight, rather than a single global rule.

Improvements

  • Offloading heuristics in ggml were reworked so ops follow their weights' backend, which can change scheduling and memory placement across mixed CPU/GPU setups.
  • DeepSeek-V4 graph construction was corrected so the model executes properly.

Bug fixes

  • Fixed DeepSeek-V4 graph issues that previously prevented correct execution.

Upgrade notes

No CLI flags or config formats changed, so this is a drop-in binary swap. Because op-to-backend assignment logic changed, re-check VRAM/RAM usage and throughput on multi-backend or partial-offload configurations (e.g. -ngl, tensor splitting) before rolling out widely. DeepSeek-V4 users should retest generation quality since graph fixes alter execution.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Adjusted the logic for offloading operations to the backend of the weights, and fixed graph issues for DeepSeek-V4.

Key Highlights

  • Adjusted offloading logic based on weight backend
  • Fixed DeepSeek-V4 graph execution

Release Notes (excerpt)

<details open> ggml : adjust logic for offloading ops to weight's backend (#25832) * ggml : adjust logic for offloading ops to weight's backend * llama : dsv4 graph fixes </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories