2026-10-22 · America/Los_Angeles · 社区动态 · #16

Massive B70 spec-decoding improvement - Intel XMX matrix engines refactored by budiony · Pull Request #29864 · ggml-org/llama.cpp

This made a huge improvement to speculate decoding speeds on my B70 running SYCL. Below examples are for Muse Glimmer Q6_K (30B) running with DFlash2. NOTE: Due to the significantly reduced cost of token verification, the n-max was increased which is why the acceptance rate is lower. Before: ``` code_python pred= 192 draft= 127 acc= 106 rate=0.835 tok/s=38.0 code_cpp pred= 192 draft= 175 acc= 113 rate=0.646 tok/s=36.2 explain_concept pred= 192 draft= 135 acc= 116 rate=0.859 tok/s=41.2 summarize pred= 188 draft= 170 acc= 142 rate=0.835 tok/s=50.1 qa_factual pred= 192 draft= 145 acc= 107…

热度 45.7 / 100；排名与评分保留该期记录。