PixelML/Qwen3.8-Flash-Next-KVA-Projector

LLKVApprox KVA projector (v2) for Qwen3.8-Flash-Next (qwen4_exp hybrid: 48 layers = 36 gated-delta-net + 12 full attention, 512-expert MoE, hyper-connections, lightning indexer). Port of kishida's LLKVApprox "Encoder-Decoder" prefill to the Flash-Next hybrid; follows the v1 Qwen3.8-27B playbook (PixelML issue #134/#139).

What it does

Prefill runs layers 0..23 over all prompt tokens; this projector predicts the approximated layers' (24..47) state fills from the boundary hyper-connection state — 18 GDN recurrence-input sets (qkv/b/a) and 6 FA (K pre-RoPE + V + lightning-indexer raw keys); layers 24..47 then run only for the last prompt position(s); decode is exact.

Training

2000 steps, cached-target distillation (zero teacher kernels in the loop), AdamW lr 1e-3, 16 x 2048-token sequences of cached teacher targets. Final held-out cosines: FA k 0.945-0.953, GDN k 0.978-0.982. Ridge ceilings (single global linear): FA k 0.62, GDN qkv 0.95 — the per-layer own-weights init + MLP heads close the FA gap.

Honest limitations

• Quality is measured as teacher-forced/greedy token agreement vs the full baseline — a diagnostic, not task accuracy.
• On this 512-expert model, bf16 GEMM differences between the CED suffix path (M=1) and the full prefill (M=T) flip near-tie router decisions; divergence is bounded but non-zero by construction (see issue #139 gate note).
• v1's suffix ladder (exact replay of the last N tokens) remains available as the quality knob.

Usage

Requires the qwen4_exp modeling code (transformers >= 5.18-dev) and the LLKVApprox engine (shipping in the club-170hx PR). Load projector.safetensors into KVAProjector(model); the frozen w0 buffers are included so no weight surgery is needed at load time.