llama.cpp
Open-Source Inference Makes Long Contexts 1.72x Faster
Page-sparse decoding boosts Qwen3-30B-A3B by 1.72x on long contexts, but cuts accuracy and highlights limits for smaller models.
1 article tagged with this topic
Page-sparse decoding boosts Qwen3-30B-A3B by 1.72x on long contexts, but cuts accuracy and highlights limits for smaller models.