llama.cpp b10603-b10615 — GLM-4.5-Air MTP, Deepseek 4 -sm tensor, Metal flash-attn vec tuning for M1 Pro/M2 Ultra/M5 Max
ggml-org
Series of llama.cpp master builds Aug 23-24. Notable: b10603 adds MTP (multi-token prediction) support for GLM-4.5-Air; b10604 adds Deepseek 4 '-sm tensor' support; b10608 fixes mtmd-video moov atom at file end; b10610 shortens virtual-device naming in CUDA and Metal; b10614 splits per-op Metal sources for parallel compile and adds kernels for col2im_1d, set_rows, Q2_0, CONV_2D_DW, fused snake activation, FWHT; b10615 adds per-device tuned (Q, NE) flash-attn vec, including FA-vec tunings for M1 Pro, M2 Ultra, and M5 Max.
Why it matters
GLM-4.5-Air MTP and Deepseek 4 -sm tensor are model-specific perf paths that matter for anyone running those architectures locally. The Metal flash-attn vec tuning for M5 Max is forward-looking — it implies Apple silicon users on the newest chips will see meaningful speedups without changing code.
Importance: 2/5
default