Red Metal dev8 — resident routed top-k layer

Dev8 extends the field-validated dev7 single-expert FFN into one correctness-first routed top-k layer.

For a selected set of experts in one routed layer:

1. the hard-bounded LRU makes every selected expert resident;
2. selected experts are protected from eviction while remaining misses are filled;
3. the native executor pins every selected Metal slot for one command buffer;
4. each expert runs gate, up, SiLU(gate) * up, and down from its resident slot;
5. a Metal kernel performs accumulator += router_weight * expert_output after each expert;
6. only the final routed-layer output is copied back to the CPU.

The first top-10 cold selection should therefore produce 10 expert loads and 30 positional GGUF reads (gate/up/down for every expert), followed by zero SSD reads during the routed-layer command buffer.

The default validation uses ten deterministic non-contiguous expert ids and positive FP32 router weights normalized to approximately 1.0. Explicit --experts and --weights lists can override these values.

Field validation#

Build and install:

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

IQ2_XS routed layer:

redlite-topk parity \
  models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
  --layer 0 \
  --top-k 10 \
  --rows 8 \
  --cache-gib 0.25

IQ1_M routed layer:

redlite-topk parity \
  models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
  --layer 6 \
  --top-k 10 \
  --rows 8 \
  --cache-gib 0.25

Success requires:

• expert loads: 10 on a cold top-10 selection;
• pread calls: 30 on that cold selection;
• SSD during top-k: 0 bytes / 0 calls;
• slots in-flight: NO after the synchronous wait;
• parity match: YES against the independent CPU reference.

This is still a correctness-first scalar path. It deliberately does not yet claim production performance. Router-network integration, real token router outputs, SIMD/simdgroup expert kernels, async prefetch, cross-layer scheduling and SSD/GPU overlap remain later milestones.