Commit Graph

Select branches

Hide Pull Requests

docs/turbo4-rematch-jun2026

experiment/asymmetric-kv

experiment/context-scaling-fix

experiment/decode-speed-parity

experiment/layer-adaptive

experiment/layer-adaptive-extended-ctx

experiment/moe-aware-gating

experiment/speed-optimization

experiment/temporal-decay

experiment/temporal-decay-prototype

main

refract-v0.1

ship/pr-90-curated

test/villines-probe-validation

#100

#100/merge

#34

#34/merge

#36

#37

#37/merge

#41

#52

#52/merge

#53

#53/merge

#61

#62

#63

#64

#65

#66

#71

#71/merge

#73

#73/merge

#75

#75/merge

#76

#76/merge

#81

#81/merge

#90

#90/merge

#91

#91/merge

#93

#94

#94/merge

#96

#96/merge

v0.3.2.1

v0.3.2.2

v0.3.2.3

Mono Color

• 5bdd8b3737 docs: move ISWA discovery to cross-model-validation.md where it belongs 3 people 2026-03-28 18:53:51 -05:00
• f2d56d60de docs: ISWA discovery story + cross-model testing methodology 3 people 2026-03-28 18:52:44 -05:00
• 3ce8c388e7 docs: cross-model complete — 7 models, Gemma FIXED, Llama 70B added 3 people 2026-03-28 18:40:54 -05:00
• 46d4e75314 docs: Gemma 2 investigation — ISWA dual-cache likely root cause, not head_dim 3 people 2026-03-28 18:30:12 -05:00
• 398cadc9e8 docs: add QC precompute result (#16) + ILP root cause analysis 3 people 2026-03-28 18:25:37 -05:00
• bcbdda285c docs: add Q·centroid precompute section to SMEM experiment doc 3 people 2026-03-28 18:14:41 -05:00
• e6fea6e363 docs: SMEM pre-dequant experiment — negative result (#15) 3 people 2026-03-28 18:09:07 -05:00
• 5ec5939750 docs: Mixtral skipped (bad GGUF), Llama 70B running 3 people 2026-03-28 18:04:33 -05:00
• e0aca7f521 docs: cross-model validation — 5 models tested, hd256 broken 3 people 2026-03-28 17:59:58 -05:00
• caee8b9080 fix: quick bench parsing — PPL field 5, decode awk pipe, NIAH table grep 3 people 2026-03-28 17:15:53 -05:00
• fd78f7e527 Add BTC donation address to README 3 people 2026-03-28 16:30:18 -05:00
• 35f3267150 feat: add --no-ref mode to quick bench for unknown models 3 people 2026-03-28 16:15:14 -05:00
• 35c5968736 Add GitHub Sponsors link to README 3 people 2026-03-28 16:10:07 -05:00
• 57baada4fc feat: add turbo-quick-bench.sh rapid smoke test 3 people 2026-03-28 15:55:41 -05:00
• f553183bf7 docs: add KLD nuance to resurrection — q4_0 lower despite worse PPL 3 people 2026-03-28 15:49:21 -05:00
• d4704f07b7 docs: NIAH 8K/100% investigation — quantization-related, not model 3 people 2026-03-28 15:24:31 -05:00
• 2b7b02824f docs: fix stale data in README + resurrection doc 3 people 2026-03-28 15:14:13 -05:00
• dde9cbda7d docs: q8_0 NIAH v2 — 30/33 (90.9%), turbo4 beats q8_0 (31/33) 3 people 2026-03-28 15:03:25 -05:00
• 2a3ce88ad9 docs: NIAH v2 turbo4 — 31/33 (93.9%), consistent edge failures 3 people 2026-03-28 14:50:26 -05:00
• 9cd29881df docs: add dense q8_0 PPL baseline (6.8884) for turbo4 comparison 3 people 2026-03-28 14:41:53 -05:00
• 74c12f2eb8 docs: dense turbo4 results — matches q8_0 decode, no regression 3 people 2026-03-28 14:40:18 -05:00
• 515354f72d docs: fix decode table column header 3 people 2026-03-28 14:21:05 -05:00
• 661bc7fb47 docs: turbo4 full characterization — PPL, prefill, decode, NIAH, KLD, PDF 3 people 2026-03-28 14:15:40 -05:00
• 36137f3c50 Merge pull request #41 from seanrasch/main TheTom 2026-03-28 11:39:02 -05:00
• c55ea45a26 docs: q4_0 coverage gaps closed — full matrix complete 3 people 2026-03-28 00:24:49 -05:00
• cb4f8e5369 docs: final polish + test suite formalized 3 people 2026-03-27 23:48:59 -05:00
• ed515e1423 docs: README final pass — ground claims, isolate core idea, soften tone 3 people 2026-03-27 23:37:57 -05:00
• 550798ae14 docs: paper final precision pass — scope claims, sharpen framing 3 people 2026-03-27 23:36:37 -05:00
• 95966a05a5 docs: README final polish — prefill parity, 50-chunk callout, soften claims 3 people 2026-03-27 23:34:49 -05:00
• 5e162be4c6 docs: q4_0 full validation + cross-format summary 3 people 2026-03-27 23:14:52 -05:00
• 419493a532 docs: README rewrite — tighter claims, proper framing, q4_0 baseline 3 people 2026-03-27 21:41:02 -05:00
• a3f4f0887c docs: add q4_0 baseline to 50-chunk 32K validation 3 people 2026-03-27 21:35:48 -05:00
• 817900cfc9 docs: 50-chunk wikitext-103 validation — strongest sparse V evidence 3 people 2026-03-27 19:53:14 -05:00
• 59efd5eddb docs: full long-context validation — direct skip rate + stronger PPL 3 people 2026-03-27 16:26:29 -05:00
• 863d23f88d docs: clarify quality claims and evaluation scope 3 people 2026-03-27 13:47:39 -05:00
• 97b071c594 docs: long-context sparse V validation — zero PPL delta through 32K 3 people 2026-03-27 13:36:35 -05:00
• b7537e5b4f docs: f16 sparse V experiment — not worth it, reverted 3 people 2026-03-27 13:13:35 -05:00
• cb55e1c145 docs: tile-level V skip experiment — not worth it, reverted 3 people 2026-03-27 12:59:42 -05:00
• 382872b3a5 docs: soften KLD comparison wording 3 people 2026-03-27 12:30:07 -05:00
• fb0aaae532 docs: KL divergence benchmarks — MoE and dense, 3 cache types each 3 people 2026-03-27 12:29:07 -05:00
• 03a2f3d69f docs: dense 27B validation — no regression, safe default 3 people 2026-03-27 12:19:21 -05:00
• 259bc337fb Merge remote-tracking branch 'upstream/main' into main Sean 2026-03-27 12:44:00 -04:00
• 4625f4e635 docs: README — sparse V is general (q8_0 validated), 3-line kernel change 3 people 2026-03-27 09:39:30 -05:00
• b81f0a4b44 docs: last two — clarity fix in 4.6, confident tone in conclusion 3 people 2026-03-27 08:20:27 -05:00
• 5ef73c7d6e docs: final cleanup — kill redundancy, tighten tone, ship it 3 people 2026-03-27 08:17:54 -05:00
• 6678911077 docs: final surgical edits — zero-cost framing, noise removal, q8_0 punch 3 people 2026-03-27 08:16:27 -05:00
• 450deaaad6 docs: six surgical edits — sharpen framing, no new content 3 people 2026-03-27 08:14:56 -05:00
• a5472c3514 docs: targeted framing updates — attention-gated computation throughline 3 people 2026-03-27 08:07:45 -05:00
• 730b32fdf1 docs: final polish — fix section numbering, sharpen generality and conclusion 3 people 2026-03-27 08:03:46 -05:00
• fcdc6e4abf docs: M2 Pro 4-mag + sparse V confirmed stacking — 0.45x → 0.73x 3 people 2026-03-27 08:01:48 -05:00
• 287a9b0914 docs: q8_0 sparse V quality validation — PPL and NIAH identical 3 people 2026-03-27 07:56:24 -05:00
• 80ab3ded47 docs: sparse V validated on q8_0 — +5% decode, confirms generality 3 people 2026-03-27 07:48:10 -05:00
• 261475eb37 docs: update README NIAH results with sparse V (9/9 perfect retrieval) 3 people 2026-03-27 07:32:50 -05:00
• 4199b6a87e docs: add link to full hardware analysis from Section 5 3 people 2026-03-27 07:24:00 -05:00
• 9bf10999e1 style: remove em dashes from sparse V paper, clean up punctuation 3 people 2026-03-27 07:22:14 -05:00
• 37772c27d8 docs: link threshold ablation logs from sparse V paper 3 people 2026-03-27 07:10:50 -05:00
• 1bc380b5c1 docs: threshold ablation sweep (τ=1e-4 to 1e-8) for sparse V paper 3 people 2026-03-27 07:07:07 -05:00
• ef05f7dd2a docs: add real-world server benchmark to sparse V paper 3 people 2026-03-27 01:51:06 -05:00
• 4d0680e70c docs: add 24K real-world PDF decode data point to README 3 people 2026-03-27 01:31:06 -05:00
• 546cb8fb48 docs: update README with sparse V decode numbers and NIAH aggregates 3 people 2026-03-27 01:17:22 -05:00
• fe7009aa0d Add technical report: Attention-Gated Value Dequantization 3 people 2026-03-27 00:33:40 -05:00
• dd73976a03 docs: M5 long-context discovery — turbo3 ceiling 28% faster than q8_0 at 32K 3 people 2026-03-27 00:11:44 -05:00
• 4a27f20956 Add quality & speed summary section to scripts README 3 people 2026-03-26 23:43:52 -05:00
• d11fd4a961 docs: update README with norm correction, 4-mag LUT, layer-adaptive, temporal decay, NIAH, community 3 people 2026-03-26 23:43:37 -05:00
• b0b9619ce3 docs: final 14-approach experiment log — complete decode optimization research 3 people 2026-03-26 23:36:51 -05:00
• 8dcccb6d8a docs: add FMA branchless + simd_shuffle results (approaches 12-13) TheTom 2026-03-26 23:26:02 -05:00
• d0036c829a docs: complete 12-approach experiment log + hardware truth + papers reviewed 3 people 2026-03-26 23:12:02 -05:00
• 081abfff05 docs: add inline block + named-reg results to scoreboard (10 approaches total) TheTom 2026-03-26 22:47:53 -05:00
• df1e37807f docs: add complete experiment scoreboard — 9 approaches, branches vs constant reads finding TheTom 2026-03-26 22:36:13 -05:00
• 8ecb090ec6 docs: decode speed hardware analysis — why M2/M1 struggle and what we did 3 people 2026-03-26 22:22:38 -05:00
• 37b31d0b52 fix: adjust turbo4 test thresholds based on empirical results SeanandClaude Opus 4.6 2026-03-26 22:37:33 -04:00
• 31e19ecb31 test: add turbo4 test coverage (14 tests) SeanandClaude Opus 4.6 2026-03-26 21:54:02 -04:00
• 9674c9bf4a fix: gitignore venv and test result dirs (accidentally committed) TheTom 2026-03-26 18:26:30 -05:00
• 94f87fe0ad docs: comprehensive experiment status update — all 2026-03-26 results 3 people 2026-03-26 18:25:59 -05:00
• deb2d8409d experiment: temporal decay prototype — 3→2 requant VIABLE 3 people 2026-03-26 17:29:58 -05:00
• d456030799 docs: add M1 Max and M2 Pro community diagnostic results 3 people 2026-03-26 17:16:57 -05:00
• b42c027917 docs: add M1 Max and M2 Pro community diagnostic results experiment/temporal-decay-prototype 3 people 2026-03-26 17:16:57 -05:00
• 22180a19f1 fix: scrub home directory from diagnostic output before zipping 3 people 2026-03-26 17:08:47 -05:00
• a3a335d4a9 docs: add AMD/ROCm experimental warning — untested, FA broken on RDNA 2 TheTom 2026-03-26 17:08:42 -05:00
• 8533410eb1 docs: add fork table for NVIDIA/AMD — point to spiritbuun's CUDA fork 3 people 2026-03-26 16:58:26 -05:00
• 2b381cc44d docs: Idea E (rotation-free outlier channeling) tested and DEAD 3 people 2026-03-26 16:51:47 -05:00
• 984ffcf79b docs: add NIAH results to README — 80-100% retrieval through 32K 3 people 2026-03-26 15:35:47 -05:00
• 8a35270b26 docs: NIAH multi-key distractor results — 100% both cache types through 32K TheTom 2026-03-26 15:33:30 -05:00
• f08e6e21c4 docs: NIAH v2 single-needle results — M5 Max baseline 3 people 2026-03-26 15:30:11 -05:00
• bee2625d9c feat: NIAH v2 — Kamradt/RULER industry-standard methodology 3 people 2026-03-26 14:56:12 -05:00
• 5c1c50ec6e fix: score from content only (not reasoning_content), 2048 max_tokens, 4K headroom TheTom 2026-03-26 14:29:16 -05:00
• bc195fe98e fix: handle Qwen3.5 thinking mode — check reasoning_content + 512 max_tokens TheTom 2026-03-26 14:02:35 -05:00
• 2d66fe96f6 fix: disable thinking mode + strip <think> tags in NIAH responses TheTom 2026-03-26 13:58:56 -05:00
• e36048f2a6 fix: add -o shorthand for --output-dir in NIAH script TheTom 2026-03-26 13:50:09 -05:00
• c8b5d7ba88 feat: NIAH benchmark — needle-in-a-haystack retrieval accuracy test 3 people 2026-03-26 13:46:21 -05:00
• ce7af8b3f9 docs: add 'plug in charger' warning — battery throttling observed on M5 Max TheTom 2026-03-26 12:26:45 -05:00
• 5cd3c65e99 fix: exclude profiles/ from setuptools package discovery (CI fix) TheTom 2026-03-26 12:10:42 -05:00
• f37f639ebd docs: add clone step to README quick start TheTom 2026-03-26 12:09:15 -05:00
• ebad090702 docs: update comparison matrix with 10 community testers from X post 3 people 2026-03-26 12:03:57 -05:00
• 5bf0aef9d1 fix: 5 Codex blockers — zip safety, exit codes, multi-GPU, PII, skip-ppl 3 people 2026-03-26 11:58:17 -05:00
• 05b9a49a9b docs: hardware comparison matrix — all tested configurations 3 people 2026-03-26 11:36:50 -05:00
• 69b33b7bf7 feat: 97% coverage (233 tests), SSD detection, comprehensive failure logging 3 people 2026-03-26 11:35:59 -05:00
• bb8a04b6a7 fix: Codex review — thread safety, crash paths, signal handling 3 people 2026-03-26 10:47:44 -05:00
• 8c9d6d4726 feat: Python diagnostic tool v5 + venv launcher + README 3 people 2026-03-26 10:43:12 -05:00
• 88af7838d0 feat: v4 — continuous background monitoring + GPU/mmap diagnostics 3 people 2026-03-26 10:06:42 -05:00