b9876

ggml-org/llama.cppb9876Jul 5, 2026by github-actions[bot]same day as b9874

What changed since b9874

low risk

b9876 adds a crash fix for tensor parallelism combined with CPU-offloaded MoE experts; no other changes noted.

Improvements

  • Split-state lookup in the meta backend now occurs before the contiguity assert, allowing mirrored non-contiguous tensors to be handled in both get_tensor and set_tensor.

Bug fixes

  • Resolves a GGML_ASSERT(ggml_is_contiguous) abort during warm-up when using -sm tensor together with -ncmoe on MoE models (issue #24886).
  • The MoE router output (ffn_moe_topk), a mirrored non-contiguous view, no longer trips the contiguity check.

Upgrade notes

No config or API changes are required; this is a drop-in replacement for b9874. Users running tensor parallelism with CPU-offloaded MoE experts should upgrade to avoid the warm-up crash.

Analyzed Oct 28, 2026

Upgrading?

Release notes summary

Fixes a crash when using tensor parallelism (-sm tensor) combined with CPU-offloaded MoE experts (-ncmoe). The fix moves the split-state lookup before the contiguity assert.

Key Highlights

  • Fixes crash for tensor parallelism with CPU-offloaded MoE experts
  • Fixes GGML_ASSERT failure for non-contiguous mirrored tensors
  • Moves split-state lookup before contiguity check
  • Fixes issue #24886

New Features

  • Stability fix for tensor parallelism with MoE experts

Release Notes (excerpt)

<details open> ggml : fix tensor-parallel + -ncmoe crash on MoE models (#25028) Tensor parallelism (-sm tensor) combined with -ncmoe (CPU-offloaded MoE experts) aborts during warm-up on MoE models with GGML_ASSERT(ggml_is_contiguous(tensor)) in ggml-backend-meta.cpp. The failing tensor is the MoE router output (ffn_moe_topk): it is mirrored (GGML_BACKEND_SPLIT_AXIS_MIRRORED, replicated across backends since routing must be identical) and happens to be a non-contiguous view.…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories