3 people af7bdf3d0c docs: complete cross-model rerun with ISWA fix + Gemma outlier log ...

Gemma 2: 3/3 NIAH all types (ISWA fix works), turbo4 slightly worse
than turbo3 (outlier, logged for investigation).
Llama 70B: confirmed 3/3 NIAH, PPL matches prior run.
Qwen MoE: sanity passed, known 8K/100% NIAH weakness confirmed.
Corrected Gemma PPL from 1-chunk estimate to proper 8-chunk values.

Co-Authored-By: tturney@psyguard.ai
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-03-28 19:15:10 -05:00

..

asymmetric_kv_blocked.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

dense_27b_sparse_v_ablation_m5.txt

docs: 50-chunk wikitext-103 validation — strongest sparse V evidence

2026-03-27 19:53:14 -05:00

dense_27b_sparse_v_clean_m5.txt

docs: dense 27B validation — no regression, safe default

2026-03-27 12:19:21 -05:00

f16_sparse_v_experiment.txt

docs: f16 sparse V experiment — not worth it, reverted

2026-03-27 13:13:35 -05:00

gemma2_investigation.txt

docs: Gemma 2 investigation — ISWA dual-cache likely root cause, not head_dim

2026-03-28 18:30:12 -05:00

gemma2_turbo4_outlier.txt

docs: complete cross-model rerun with ISWA fix + Gemma outlier log

2026-03-28 19:15:10 -05:00

long_context_ppl_full_validation.txt

docs: full long-context validation — direct skip rate + stronger PPL

2026-03-27 16:26:29 -05:00

long_context_ppl_validation.txt

docs: clarify quality claims and evaluation scope

2026-03-27 13:47:39 -05:00

m2_pro_4mag_sparse_v.txt

docs: M2 Pro 4-mag + sparse V confirmed stacking — 0.45x → 0.73x

2026-03-27 08:01:48 -05:00

niah_8k_100pct_investigation.txt

docs: NIAH 8K/100% investigation — quantization-related, not model

2026-03-28 15:24:31 -05:00

q4_0_coverage_gaps_closed.txt

docs: q4_0 coverage gaps closed — full matrix complete

2026-03-28 00:24:49 -05:00

q4_0_full_validation.txt

docs: q4_0 full validation + cross-format summary

2026-03-27 23:21:45 -05:00

q8_0_sparse_v_ablation_m5.txt

docs: sparse V validated on q8_0 — +5% decode, confirms generality

2026-03-27 07:48:10 -05:00

q8_0_sparse_v_quality_m5.txt

docs: q8_0 sparse V quality validation — PPL and NIAH identical

2026-03-27 07:56:24 -05:00

skip_rate_direct_measurement.json

docs: full long-context validation — direct skip rate + stronger PPL

2026-03-27 16:26:29 -05:00

tau_1e-4.txt

docs: threshold ablation sweep (τ=1e-4 to 1e-8) for sparse V paper

2026-03-27 07:07:07 -05:00

tau_1e-5.txt

docs: threshold ablation sweep (τ=1e-4 to 1e-8) for sparse V paper

2026-03-27 07:07:07 -05:00

tau_1e-6.txt

docs: threshold ablation sweep (τ=1e-4 to 1e-8) for sparse V paper

2026-03-27 07:07:07 -05:00

tau_1e-7.txt

docs: threshold ablation sweep (τ=1e-4 to 1e-8) for sparse V paper

2026-03-27 07:07:07 -05:00

tau_1e-8.txt

docs: threshold ablation sweep (τ=1e-4 to 1e-8) for sparse V paper

2026-03-27 07:07:07 -05:00

tile_skip_experiment.txt

docs: tile-level V skip experiment — not worth it, reverted

2026-03-27 12:59:42 -05:00

turbo4_2plus1_packing_fix.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_4bit_long_context.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_4bit_polarquant_results.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_component_isolation.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_cuda_buun_data.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_dequant_then_attend_wip.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_direct_extract_results.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_full_characterization.txt

docs: NIAH v2 turbo4 — 31/33 (93.9%), consistent edge failures

2026-03-28 14:50:26 -05:00

turbo4_prefill_comparison.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_qjl_ablation.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_setrows_fix_results.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

turbo4_vs_buun_comparison.txt

docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF

2026-03-28 14:19:35 -05:00

wikitext103_32k_50chunk_q4_0_validation.txt

docs: add q4_0 baseline to 50-chunk 32K validation

2026-03-27 21:35:48 -05:00

wikitext103_32k_50chunk_validation.txt

docs: 50-chunk wikitext-103 validation — strongest sparse V evidence

2026-03-27 19:53:14 -05:00