Red Lite dev17 — audited 48-layer decoder stack

Target: Apple M4 Pro, 24 GiB unified memory, Bartowski Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf.

Dev17 begins after independent CPU/Metal parity passed for both complete decoder block variants. This checkpoint extracts those validators into reusable native C entry points and propagates separate CPU and Metal hidden vectors through the complete transformer trunk.

Structural dispatch#

redlite-decoder-stack does not infer the hybrid pattern from a fixed interval. It scans the real GGUF tensor directory and requires:

a contiguous layer sequence;
both attention norm tensors on every layer;
exactly one recurrent or full-attention structural signature per layer;
no mixed or unknown layers.

The target model reports 48 layers: 36 recurrent DeltaNet blocks and 12 full-attention blocks at layers 3,7,11,15,19,23,27,31,35,39,43,47.

Execution boundary#

For one deterministic hidden vector, the diagnostic executes each audited block in order:

recurrent DeltaNet or GQA full attention;
attention residual;
post-attention RMSNorm;
full-width top-10 routed MoE plus shared expert;
final residual;
pass the complete 2048-value output to the next layer.

CPU and Metal outputs are propagated independently. Isolated block CLIs retain their strict same-input substage checks; stack execution evaluates cumulative block parity so normal upstream floating-point error is not misreported as a local RMSNorm failure. A top-k router ID divergence remains a hard failure.

Field validation#

make native

.deps/redmetal/redlite-decoder-stack parity \
  models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
  --position 7 \
  --top-k 10 \
  --cache-mib 256

Observed on the target M4 Pro:

GGUF layer map       : 48 layers / recurrent=36 full-attention=12
dispatch totals      : recurrent=36 full-attention=12
stack max abs/rel    : 0.000238419 / 0.133136
elapsed              : 9049.723 ms
COMPLETE 48-LAYER DECODER STACK: YES

The large maximum relative error is attached to values near zero; the maximum absolute difference over every 2048-wide layer output is 2.38419e-04.

Scope boundary#

This is a correctness-first, single-token transformer-trunk diagnostic. Previous recurrent/KV states are deterministic parity fixtures, and model structures are reopened per block. It is not yet a persistent multi-token runtime and the reported elapsed time is not a throughput benchmark.

The next correctness gates are input embedding, final RMSNorm and LM-head logits, followed by tokenizer/sampling and persistent multi-token state management.