BriefPulse Practical AI · Working notes on AI you can actually use. RSS · BriefPulse network
BriefPulse Practical AI

What changed in AI, what it is useful for, and what you can do with it.

16 September 2026

Brief

llama.cpp b11009 fixes tensor-parallel split states for fused QKV in gemma4 and qwen35

Build b11009 of llama.cpp carries a tensor-parallel fix for fused QKV in gemma4 and qwen35. The release notes give one short changelog line plus a model-side explanation, so this brief sticks to what that line states.

The change is described as a fix to split state and granularity for fused QKV in gemma4 and qwen35. The model-side note says split states for attn_qkv should be calculated from n_head * n_embd_head_k, which is required for gemma4 with --fuse-qkv, where n_embd is 5376 but Q is 8192.

The same build adds handling for fused full attention layers in qwen35/qwen35moe, and carries a TODO tagged for split qgate in qwen. Prebuilt binaries are listed across macOS, Linux, Android, Windows and openEuler, with the macOS Apple Silicon KleidiAI build and openEuler flagged as disabled.

Our reading

For anyone running llama.cpp across multiple devices with --fuse-qkv, the split calculation is the part that decides how work is divided, so a mismatch between n_embd and the Q dimension is the kind of detail that matters when shuffling these models across GPUs. This is narrow, model-specific work rather than a general speed or quality change, and it is most relevant to people already serving gem…

What to do or watch

Before switching a tensor-parallel gemma4 or qwen35 deployment to b11009, verify the --fuse-qkv path for the 5376/8192 dimension case on your own hardware, and note that the split qgate item for qwen remains an open TODO in this build.

Source details and supporting facts

Each line is stated by the page named above it.

Stated by github.com

  • The changelog line reads: "TP: fix split state and granularity for fused QKV gemma4, qwen35 ( #28965 )".
  • Split states for attn_qkv are calculated from n_head * n_embd_head_k, required for gemma4 with --fuse-qkv, where n_embd is 5376 but Q is 8192.
  • The build adds handling for fused full attention layers for qwen35/qwen35moe.
  • The build adds a TODO tagged [TAG_SPLIT_QGATE_QWEN].
  • Prebuilt binaries are listed for macOS, Linux, Android, Windows and openEuler targets, with macOS Apple Silicon (arm64, KleidiAI enabled) marked DISABLED.

Sources

  1. llama.cpp releases (GitHub)Text stored 16 September 2026

How this story was checked. Written from the 1 page listed above, stored 16 September 2026; claims checked against that stored text on 16 September 2026.

What that means
  • 5 of 6 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
  • Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
  • The check reads stored text only: no claim rests on a fresh look that did not happen.
  • Where the reporting was silent, the text says so instead of filling the gap.

More from Practical AI