llama: create new fused op for QSA mask by am17an · Pull Request #30640 · ggml-org/llama.cpp
it's time to update your llama.cpp for another speedup (sparse attention models)
Proof1 community thread
Sources (1)
- [1]llama: create new fused op for QSA mask by am17an · Pull Request #30640 · ggml-org/llama.cppr/LocalLLaMA (top, daily) · Oct 21, 07:34 PM
it's time to update your llama.cpp for another speedup (sparse attention models)
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026Mica v0.5: The smartest 4B decision model is also the fastest one. 20 ms per answer, 25x faster than Jev
- Oct 21, 2026Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
- Oct 20, 2026Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0
- Oct 20, 2026How I built my ‘everything’ app
- Oct 20, 2026Voxtral 3
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
