PulseAugur
EN
LIVE 16:52:08

llama.cpp adds MMA FlashAttention support for Gemma and Qwen models

By PulseAugur Editorial · [1 sources] ·

The llama.cpp project has released an update, b12285, which introduces support for MMA FlashAttention with head sizes of 40 and 72. This enhancement is particularly relevant for vision models like Gemma and Qwen, which utilize these specific head sizes. The update enables tensor-core dispatch for these sizes on NVIDIA Turing GPUs and newer, while older architectures like Volta and AMD GPUs will continue to use the tile kernel. AI

IMPACT Improves inference performance for specific vision models on compatible NVIDIA hardware.

RANK_REASON This is a software update for a specific project (llama.cpp) that improves performance for certain model architectures, rather than a new model release or significant industry-wide development.

Read on llama.cpp — Releases →

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

llama.cpp adds MMA FlashAttention support for Gemma and Qwen models

How we ranked this

Signal score
5 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
This is a software update for a specific project (llama.cpp) that improves performance for certain model architectures, rather than a new model release or significant industry-wide development.
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, model release
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
Same-day
Cluster formed today. Ranking reflects the current source set at time of score.

Full methodology in our editorial standards.

COVERAGE [1]

  1. llama.cpp — Releases TIER_1 English(EN) · sihanyu03 ·

    b12285: cuda : fattn mma head sizes 40 72 (#30285)

    <ul> <li>cuda : add MMA FlashAttention support for head sizes 40 and 72</li> </ul> <p>Head sizes 40 and 72 (used e.g. in vision towers of Gemma 3/4 and Qwen3-VL) were excluded from every tensor-core dispatch branch and fell back to the tile kernel.</p> <p>The fp16 MMA fragment co…

RELATED ENTITIES

RELATED TOPICS