Red Metal dev7 — resident single-expert FFN

Dev7 moves the field-validated IQ2_XS and IQ1_M arithmetic into a hard-bounded resident expert cache.

For one (layer, expert) miss, gate/up/down are read once from the GGUF directly into one shared Metal slot. The same resident slot is then used for:

gate matvec
up matvec
SiLU(gate) * up
down matvec

The validation probe intentionally compares only a small selectable set of final down rows, but both gate and up projections are evaluated in full so the intermediate FFN vector is real.

The native pool enforces a hard byte budget, slab allocation, global Python-side LRU recycling and native in-flight slot protection. expert-ffn-parity also asserts that the Metal execution phase performs zero additional GGUF reads after residency.

Field commands#

Build and install:

git pull --ff-only
./scripts/install.sh
make redmetal
make test

IQ2_XS layer:

redlite-ffn parity \
  models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
  --layer 0 --expert 0 --rows 8 --cache-gib 0.25

IQ1_M layer:

redlite-ffn parity \
  models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
  --layer 6 --expert 0 --rows 8 --cache-gib 0.25

Success requires parity match: YES, SSD during FFN: 0 bytes, non-zero gate/up/down GPU addresses and slot in-flight: NO after the synchronous wait.

This remains a correctness-first scalar Metal path. SIMD-group kernels, selected-expert batching, router weights, top-k accumulation, asynchronous SSD prefetch and overlap are later milestones.