v1.4.0 — Asymmetric NF4: one robust KV-cache codebook

Asymmetric (zero-point) NF4 — key_nf4_asym / TurboQuantKVCache.robust()

Symmetric NF4 scales each key channel by abs-max about zero, but KV keys carry a large per-channel DC offset — so a symmetric grid wastes ~half its codes. On high-ratio-GQA models this silently collapses generation:

Qwen2.5-7B (GQA 7:1)	fp16	NF4	asym-NF4
LongBench qasper	43.8	4.7 💥	41.9
WikiText-2 ppl	7.46	74.7	7.50

key_nf4_asym subtracts the per-channel mean, NF4-quantizes the residual, and adds it back — one calibration-free codebook that's near-fp16 on every architecture tested: it ties symmetric NF4 where NF4 already works (Llama-2-7B/13B, Mistral-7B) and rescues the collapse where it doesn't, at no extra bit cost.

cache = TurboQuantKVCache.robust(head_dim=128, n_heads=32)   # asym-NF4 + 2% outliers, 4-bit K/V

Also in this release

• robust() factory — the recommended calibration-free recipe.
• Compact NF4/asym-NF4 storage — per-channel scalars instead of the expanded level table; compression ratio 6.3× → 7.9×.
• Cross-model matrix (benchmarks/kvquant_matrix/): 4 models × {fp16/NF4/uniform/asym-NF4}, LongBench + WikiText-2, a decision guide, and a TMLR paper draft.
• ty (Astral type checker) added to CI as an advisory step.

Honest limitation

All 4-bit KV quant — asym-NF4 included — degrades on very-long-generation tasks (e.g. 512-token summarization) as the small residual key error compounds across the decode. This is generation-length-driven and affects MHA and GQA models alike (not a codebook issue).

Full notes: CHANGELOG.md.