v0.2.3

Release Notes — v0.2.3

New model support

• Added inference support for MiniCPM text models:

  • MiniCPM5-2B
  • MiniCPM-V-4.7-35B-A3B (MoE)

  Both models have been verified on Spacemit K1 and K3 boards with the new
  ggml-spacemit device backend. Reference numbers on K3 (llama-bench, pp128/tg128,
  new backend): MiniCPM5-2B ~178 / ~17.8 t/s, MiniCPM-V-4.7-35B-A3B ~43 / ~8.6 t/s.

GDN operator optimization

• Optimized the GDN (Gated DeltaNet) kernels in the ggml-spacemit backend:

  • The fused GDN row kernel now uses an m4-width vector path when the state size
    fits, avoiding register spills in the row loop.
  • Added a fast path for GDN convolution-state copies (strided vlsseg3e32/vsseg3e32),
    and hoisted overlap/disjoint checks out of the inner loop.
  • Added an early return for empty ranges in get_rows.

  This improves decode performance for GDN-architecture models (Qwen3.5 family).

Notes for upgrading

• -mmp / --no-mmap were replaced by the upstream -lm load-mode option
  (default AUTO; use -lm none to disable mmap).