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

• 693a391333 fix: address Codex + Gemini review findings 3 people 2026-03-26 09:34:29 -05:00
• 0e235113ff feat: hardware replay system + M5 Max baseline profile 3 people 2026-03-26 09:15:29 -05:00
• 31be9fb167 feat: comprehensive hardware diagnostic script + 71 tests 3 people 2026-03-26 09:03:59 -05:00
• 0b95e6456e docs: layer-adaptive mode 2 verified 2K-32K — q8_0 quality at 3.5x compression 3 people 2026-03-26 00:17:35 -05:00
• 48115e1cb7 Merge branch 'main' into experiment/layer-adaptive-extended-ctx TheTom 2026-03-25 23:37:29 -05:00
• 2461b261d1 docs: abandon decode-speed-parity branch — 11 experiments exhausted experiment/decode-speed-parity 3 people 2026-03-25 23:32:42 -05:00
• 71742c0e83 docs: experiments 10-11 dead — packed uint4 -5.5%, fused FA +0.9% 3 people 2026-03-25 23:30:28 -05:00
• 9d4c0c4a5a docs: experiment 9 (cn[] precompute) — DEAD END, -10.5% decode regression 3 people 2026-03-25 23:21:54 -05:00
• 47df1ebd46 docs: update decode numbers — 0.92x short, 0.88x 8K, 0.72x 48K 3 people 2026-03-25 23:08:41 -05:00
• fdedbfe039 docs: variant A (half LUT + float norm) is new best — +8.7% at 48K 3 people 2026-03-25 22:51:44 -05:00
• dd440be4b8 docs: add decode speed table with context scaling to README 3 people 2026-03-25 22:43:51 -05:00
• 8b4eb0a173 docs: full test results for fp16 LUT — net positive, recommend merge 3 people 2026-03-25 22:23:19 -05:00
• 7fa354b7e8 docs: Mario PDF test confirms decode scaling — 0.66x q8_0 at 48K context 3 people 2026-03-25 22:08:23 -05:00
• c11cd3f8ee ci: real-world PDF benchmark script (Mario's 70-page fitness PDF test) 3 people 2026-03-25 22:02:33 -05:00
• b525664e3c docs: complete decode investigation — LUT cache thrashing scales with context 3 people 2026-03-25 22:01:40 -05:00
• 0c090a7f2f data: decode ratio degrades with context depth (0.88x@12tok → 0.76x@8K) 3 people 2026-03-25 21:49:43 -05:00
• 8c9e7a1514 docs: decode speed status report — 0.84x q8_0, experiment branch was net negative 3 people 2026-03-25 21:34:05 -05:00
• 55f8f5aa77 docs: decode speed investigation — LUT is the entire 16% gap 3 people 2026-03-25 21:09:11 -05:00
• 64209678ab docs: extended benchmarks + external validation (Mario M1 Max 32K) experiment/layer-adaptive-extended-ctx 3 people 2026-03-25 20:59:43 -05:00
• b4fa8b6810 docs: TurboQuant+ experiments master document — status, findings, dead ends 3 people 2026-03-25 20:49:49 -05:00
• d502cd41de Merge branch 'main' into experiment/asymmetric-kv experiment/asymmetric-kv TheTom 2026-03-25 20:27:22 -05:00
• 7b3124abfd Merge branch 'main' into experiment/moe-aware-gating experiment/moe-aware-gating TheTom 2026-03-25 20:27:22 -05:00
• 4ded8b89aa Merge branch 'main' into experiment/temporal-decay experiment/temporal-decay TheTom 2026-03-25 20:27:22 -05:00
• 9208524a8b Merge branch 'main' into experiment/layer-adaptive experiment/layer-adaptive TheTom 2026-03-25 20:27:22 -05:00
• 367bb2dbbb docs: update README with context scaling data + gate script info experiment/speed-optimization experiment/context-scaling-fix 3 people 2026-03-25 20:25:24 -05:00
• eed5a6310e ci: quality+speed gate — checks PPL + context scaling before push 3 people 2026-03-25 20:22:52 -05:00
• 9be9ca18c1 docs: extended context scaling verified — 0.99x q8_0 through 32K context 3 people 2026-03-25 20:21:22 -05:00
• 034d4c911e docs: context scaling FIXED — optimized dequant gives flat 98% of q8_0 3 people 2026-03-25 20:13:40 -05:00
• e5fcbe34db docs: group-32 rotation failed (PPL 7.06) 3 people 2026-03-25 19:31:11 -05:00
• 0255ba2d7f docs: context scaling fix investigation — group-32 rotation approach 3 people 2026-03-25 19:22:06 -05:00
• b96d0d16f5 docs: MoE-aware gating — hypothesis invalidated 3 people 2026-03-25 18:50:00 -05:00
• 4b7c5e60e1 docs: temporal decay design — 2-tier MVP needs turbo2 implementation 3 people 2026-03-25 18:50:00 -05:00
• c75d79d63a docs: layer-adaptive results — last 8 layers at q8_0 matches q8_0 quality at 3.5x compression 3 people 2026-03-25 18:44:37 -05:00
• d9e1090f1d docs: asymmetric K/V experiment results — needs FA template work 3 people 2026-03-25 18:41:07 -05:00
• 6406922ace docs: update all docs with q8_0 parity results (2747 tok/s, 4.6x compression) 3 people 2026-03-25 18:01:14 -05:00
• d50ef98f5b docs: block-32 breakthrough — q8_0 parity at 4.6x compression 3 people 2026-03-25 17:34:27 -05:00
• 1bd4fcb181 docs: experiment 3 results — graph WHT gives 2095 tok/s (0.78x q8_0) 3 people 2026-03-25 17:23:25 -05:00
• d06ff09f1c docs: experiment 1 results — half4 WHT gives 1411 tok/s (+31%) 3 people 2026-03-25 17:00:58 -05:00
• 445a3dc4c7 docs: move engineering docs here from llama.cpp fork 3 people 2026-03-25 16:16:58 -05:00
• 2d592a6f30 docs: link engineering docs from llama.cpp fork in README 3 people 2026-03-25 16:15:54 -05:00
• bd2843e67d docs: update speed table with fp16 WHT results (1074 tok/s, 0.40x q8_0) 3 people 2026-03-25 16:12:32 -05:00
• 4c67c02340 fix: honest README — PPL 6.19 confirmed, speed claims corrected 3 people 2026-03-25 14:17:34 -05:00
• bd3b4036c3 feat: near-parity with q8_0 — 91% MoE, 97% Qwopus at 4.6× compression 3 people 2026-03-25 11:58:51 -05:00
• b8c4650c03 feat: final v1 benchmarks — 73-88% of q8_0 speed at 4.9× compression 3 people 2026-03-25 11:35:59 -05:00
• b63c5aca98 feat: update benchmarks — 51.4 tok/s MoE, 14.6 tok/s Qwopus with pre-rotate-queries 3 people 2026-03-25 10:31:11 -05:00
• 128809e426 docs: update README with speed ceiling data and pre-rotate-queries plan 3 people 2026-03-25 10:00:39 -05:00
• b483a097b3 fix: update benchmarks with real Metal numbers (was CPU fallback) #23 3 people 2026-03-25 09:16:29 -05:00
• af40f0182a docs: update README with correct llama.cpp fork URL 3 people 2026-03-25 08:28:32 -05:00
• 43bdecc76b docs: comprehensive README rewrite for public release 3 people 2026-03-25 07:58:27 -05:00
• b28a2319db feat: add benchmark results and updated roadmap to README 3 people 2026-03-25 07:54:28 -05:00
• 7e469a0dc9 feat: benchmark results — turbo3/turbo4 vs q8_0/q4_0 on M5 Max #22 3 people 2026-03-25 01:06:59 -05:00
• 45586131b0 fix: remove CLAUDE.md and dev workflow from public repo 3 people 2026-03-25 00:22:22 -05:00
• 6634deebd0 feat: README overhaul, Apache 2.0 license, benchmark script #22 3 people 2026-03-25 00:09:59 -05:00
• 338e978b09 feat: validate rotation Gaussianization on real KV tensors #18 #20 3 people 2026-03-24 21:36:34 -05:00
• 4e8443f24e fix: real model validation — DynamicCache API, Phase A results #18 3 people 2026-03-24 21:34:03 -05:00
• c1a10e4a64 feat: input validation, README, real model validation script #12 #18 3 people 2026-03-24 21:29:24 -05:00
• 1fa6d79b10 feat: outlier comparison benchmark — matches Prince Canuma's ratios 3 people 2026-03-24 21:21:53 -05:00
• d0e70cd24f feat: outlier channel strategy for 2.5-bit and 3.5-bit quantization #17 3 people 2026-03-24 21:20:55 -05:00
• c3aad548ff test: achieve 100% code coverage across all modules 3 people 2026-03-24 21:18:46 -05:00
• f1c2397680 feat: integration test at real Qwen 3.5 KV cache dimensions #16 3 people 2026-03-24 21:15:14 -05:00
• dccd96bc71 fix: add norm extraction to PolarQuant — critical for real KV tensors #15 3 people 2026-03-24 21:11:22 -05:00
• 361ea4cab0 feat: add demo benchmark script for quick validation 3 people 2026-03-24 21:05:53 -05:00
• 8ec1a07bfb feat: TurboQuant, KV cache compressor, utils, distortion suite #6 #7 #8 #10 3 people 2026-03-24 21:04:56 -05:00
• 4d34f5148c feat: PolarQuant quantizer + QJL 1-bit quantizer #4 #5 3 people 2026-03-24 21:02:06 -05:00
• fe7c5f9ad2 feat: codebook construction for PolarQuant (codebook.py) #3 3 people 2026-03-24 21:00:07 -05:00
• e680643d7c fix: address roast review findings on scaffold #1 3 people 2026-03-24 20:55:30 -05:00
• 0e642a883c feat: random rotation matrix generation (rotation.py) #2 3 people 2026-03-24 20:53:52 -05:00
• 8fa1d293f9 infra: project scaffold — CI, pyproject, CLAUDE.md, PLAN.md #1 3 people 2026-03-24 20:43:16 -05:00