fix(memory): run cleanly on 32 GB Macs + Mac mini M6 benchmarks - #173

Merged

solderzzc merged 7 commits into

mainfrom

fix/fullgpu-cache-limit

Sep 24, 2026

Commits on Sep 24, 2026

fix(memory): cap the MLX cache on the full-GPU strategy; add a memory-guarded bench harness

authored and
simba
committed
3457c05
fix(memory): KV estimate counts only layers that hold a KV cache

authored and
simba
committed
3d2b64a
fix(load): fall back to text-only when an auto-detected VLM fails to load

authored and
simba
committed
396c960
docs: Mac mini M6 (base, 32 GB) benchmarks, recordings and demo client

authored and
simba
committed
5eecf22
fix(load): only fall back to text-only for VLM checkpoint mismatches

authored and
simba
committed
398252e
docs(m6): re-measure Gemma 4 and Qwen3.8 on main (mlx-swift-lm 460ff81) + this PR

authored and
simba
committed
7773c39
Merge remote-tracking branch 'origin/main' into fix/fullgpu-cache-limit
solderzzc
committed
a17b094