Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(tune): report Metal backend in plan label on Apple Silicon
JustVugg75e3fb15887e
diff --git a/c/coli b/c/coliindex 7ddbbfdd..b04c9cae 100755--- a/c/coli+++ b/c/coli@@ -619,7 +619,17 @@ def env_for(a):print(f" {C.dim}[TUNE] applied measured profile · +{gain:.1f}% calibration throughput{C.r}",file=sys.stderr)rt=plan["tiers"]["ram"]; vt=plan["tiers"]["vram"]- gpu=f" · VRAM {format_bytes(vt['budget_bytes'])}" if has_cuda and vt["devices"] else " · CPU"+ # The plan label used to derive the backend from CUDA alone, so on+ # Apple Silicon it printed " · CPU" even when the calibration child+ # processes ran on Metal (COLI_METAL is inherited through env_for and+ # environment_for_plan, so the measurements are real). Report the+ # active backend: CUDA VRAM first, then Metal, then CPU.+ if has_cuda and vt["devices"]:+ gpu=f" · VRAM {format_bytes(vt['budget_bytes'])}"+ elif str(e.get("COLI_METAL","")).strip() not in ("","0"):+ gpu=" · Metal (unified memory)"+ else:+ gpu=" · CPU"print(f" {C.dim}[PLAN] RAM {format_bytes(rt['budget_bytes'])} · cap {rt['cache_slots_per_layer']}/layer{gpu}{C.r}",file=sys.stderr)else:# Windows: a bare `coli chat` (no --gpu/--vram/--auto-tier) used to ALWAYS