PulseAugur
EN
LIVE 21:22:15
(CA) b11866: metal : remove mul_mm_id amax rescaling (#30924)

llama.cpp optimizes Metal performance with kernel changes

By PulseAugur Editorial · [1 sources] ·

The llama.cpp project has released version b11866, which includes optimizations for Metal, specifically removing amax rescaling in the mul_mm_id kernel. This change is intended to keep MoE down-projection activations within f16 range before operations run. The release also drops two-stage reduction kernels and associated dispatch, with assisted testing noted for Qwen3.8-Flash-Next-GGUF on a MacBook with an M5 Max chip. AI

IMPACT Performance optimizations for local LLM inference on Apple hardware.

RANK_REASON This is a software update/optimization for a specific tool, not a frontier release, significant industry move, or academic research.

Read on llama.cpp — Releases →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

llama.cpp optimizes Metal performance with kernel changes

How we ranked this

Signal score
7 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
This is a software update/optimization for a specific tool, not a frontier release, significant industry move, or academic research.
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. llama.cpp — Releases TIER_1 (CA) · ggerganov ·

    b11866: metal: remove mul_mm_id amax rescaling (#30924)

    <p>The src1 power-of-two rescale in kernel_mul_mm_id is superseded by the<br /> per expert-token amax scaling done in the compute graph<br /> (ref: <a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30861">#30861</a>), which keeps the<br /> MoE d…

RELATED ENTITIES

RELATED TOPICS