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

• 3aac6b545f docs: update MLX Quick Start — only TheTom/mlx needed, stock mlx-lm works TheTomandClaude Opus 4.6 2026-04-05 16:52:13 -05:00
• ca73182421 docs: update MLX headline to 96% decode TheTomandClaude Opus 4.6 2026-04-05 14:51:07 -05:00
• e33ff8c5f5 docs: update 35B to 96% decode with fused+boundary (was 56%) TheTomandClaude Opus 4.6 2026-04-05 14:32:39 -05:00
• b424ad85f1 docs: correct MLX headline to 81% with clean GPU benchmark TheTomandClaude Opus 4.6 2026-04-05 14:18:12 -05:00
• e51e71e1bf fix: add boundary layer protection to MLX quality suite TheTomandClaude Opus 4.6 2026-04-05 13:19:10 -05:00
• 276d4e65fe docs: add Quick Start, make_turbo_cache API, update status to 93% TheTomandClaude Opus 4.6 2026-04-05 13:17:30 -05:00
• 89c9e9d9c2 docs: update MLX headline to 93% decode with fused + boundary fix TheTomandClaude Opus 4.6 2026-04-05 12:24:42 -05:00
• b7d3e51879 docs: add M2, context scaling, MLX vs llama.cpp comparison to README TheTomandClaude Opus 4.6 2026-04-05 11:13:18 -05:00
• e1d4ff8d23 docs: add NIAH/KLD quality validation — asymmetric mandatory for dense models TheTomandClaude Opus 4.6 2026-04-05 09:32:30 -05:00
• 40a8bf46e5 Fix MLX quality suite: hybrid model support, KLD compression threshold 3 people 2026-04-05 09:05:34 -05:00
• fe022740c5 Add MLX quality suite for NIAH, KLD, and context scaling tests 3 people 2026-04-05 09:02:16 -05:00
• f165218762 docs: add 27B dense + 7B/phi-4 MLX results to README TheTomandClaude Opus 4.6 2026-04-05 08:32:40 -05:00
• a2dce218a3 docs: expand MLX results with prefill, PPL, 35B data, fix branch link TheTomandClaude Opus 4.6 2026-04-05 07:49:58 -05:00
• fbba659d07 docs: fix MLX branch name to feature/turboquant-plus TheTomandClaude Opus 4.6 2026-04-05 07:47:04 -05:00
• cc10343bc2 docs: add experimental MLX framework port section to README TheTomandClaude Opus 4.6 2026-04-05 07:19:42 -05:00
• 1d4de7be79 docs: add WaveboSF fused mmvq kernel benchmarks (signalnine V12) 3 people 2026-04-04 14:47:31 -05:00
• 1dc7113017 docs: add primoco RoPE root cause + 100% TBQ accuracy at native ctx_size 3 people 2026-04-04 14:46:23 -05:00
• cf634ec796 docs: add redwolfweb head_dim=256 fix confirmation (RTX 5090, all 6 K/V combos) 3 people 2026-04-04 14:45:10 -05:00
• a76b1d79ad docs: add AmesianX deterministic task benchmark (DGX Spark, Qwen3-14B) 3 people 2026-04-04 14:43:24 -05:00
• 3ff9bf96cd docs: add varjoranta vLLM V precision fix as external validation 3 people 2026-04-04 14:41:54 -05:00
• a7f559777c docs: add WaveboSF RTX 4090 + RTX 5090 benchmark data 3 people 2026-04-04 09:57:44 -05:00
• 8ce57b8bc8 docs: restore Gemma 4 31B row — different model from 26B A4B MoE 3 people 2026-04-04 08:11:24 -05:00
• 250ed960db docs: add Gemma 4 weight compression results (RTX 5090, theaaronhughes) 3 people 2026-04-04 08:09:41 -05:00
• abcda965a2 scripts: add mlx_lora_to_gguf.sh for MLX LoRA → GGUF conversion 3 people 2026-04-03 22:11:50 -05:00
• 1703deec79 docs: add Gemma 4 KV (5090), APEX-I-Quality (3050), PRO 6000 stacking results 3 people 2026-04-03 21:14:48 -05:00
• b0ee4aa675 docs: add V100 and GTX 1080 Ti to community results 3 people 2026-04-03 20:24:44 -05:00
• b18fdb4669 docs: add nihilistau Vilenkin spectral analysis (preliminary, needs cross-arch validation) 3 people 2026-04-03 16:03:43 -05:00
• fa04fd8101 docs: add RX 9070 XT (RDNA 4) weight compression results 3 people 2026-04-03 15:11:20 -05:00
• 27f335c699 docs: add adrianosousa M4 Pro independent validation (asymmetric paper) 3 people 2026-04-03 14:21:46 -05:00
• 748c9efd72 docs: add RTX PRO 6000 Blackwell Workstation results (Config I +8.5% faster than Q8_0) 3 people 2026-04-03 13:30:55 -05:00
• 77b2e1b0c2 docs: add vLLM PR #38479 as independent validation (asymmetric + boundary papers) 3 people 2026-04-03 13:26:42 -05:00
• 7001bde828 docs: update all references — PR #45 merged to main 3 people 2026-04-03 11:27:00 -05:00
• 8cdc958062 docs: note X/Twitter as additional data source for community results 3 people 2026-04-03 10:37:19 -05:00
• 5b01015489 docs: add independent validation to weight compression paper 3 people 2026-04-03 10:33:43 -05:00
• cba570dfd0 docs: note living document, updated as findings come in 3 people 2026-04-03 10:28:46 -05:00
• d1c81e5cbe docs: expand independent validation section with full data tables 3 people 2026-04-03 10:26:21 -05:00
• ae7f05cc8e docs: add community results page for TQ4_1S weight compression 3 people 2026-04-03 10:20:38 -05:00
• e8d192e418 docs: add signalnine CUDA optimization log to references, update CUDA decode analysis 3 people 2026-04-02 15:10:28 -05:00
• 71fecb3aae docs: point test results to PR #45, update backend support (Metal + CUDA + AMD HIP) 3 people 2026-04-02 11:09:09 -05:00
• 3cba8d2601 docs: add AMD HIP Windows build instructions with required optimization flags 3 people 2026-04-02 11:02:27 -05:00
• 4dfa32aad8 Add GitHub Sponsors funding link 3 people 2026-04-02 10:41:15 -05:00
• 7d415ab826 docs: update David Y. Tan attribution per his request 3 people 2026-04-02 10:36:38 -05:00
• c94aa4ef9c docs: Sjoerd asymmetric correction + MTMD tip + community update 3 people 2026-04-02 10:07:30 -05:00
• 826f926cc1 docs: remove duplicate quantize steps, reference Quick Test instead 3 people 2026-04-02 10:06:40 -05:00
• ff45fe41d7 docs: update weight compression paper — Metal + CUDA, not Metal only 3 people 2026-04-02 10:05:34 -05:00
• 8ef30366df docs: include filenames in size comparison output 3 people 2026-04-02 10:03:26 -05:00
• e1c0c69211 docs: show model size before/after compression in quick test 3 people 2026-04-02 10:01:38 -05:00
• 0a5a151c37 docs: add quick test workflow for PR #45 testers 3 people 2026-04-02 10:00:55 -05:00
• ce4b3b0592 docs: move PR #45 link to top of weight compression paper 3 people 2026-04-01 21:39:07 -05:00
• f15e998190 docs: add PR #45 link to paper and getting-started 3 people 2026-04-01 21:25:14 -05:00
• d860058da9 docs: TQ4_1S weight compression paper and getting-started update 3 people 2026-04-01 21:13:36 -05:00
• ec29f21fbf Add Sjoerd L40S validation, KLD context-length stability data 3 people 2026-04-01 15:09:08 -05:00
• 05fde211c0 Add head_dim=64 limitation warning to getting-started guide 3 people 2026-04-01 12:17:43 -05:00
• e93a0027e1 Correct turbo4-K bug attribution in asymmetric paper 3 people 2026-04-01 11:42:25 -05:00
• c760dc58d1 Update docs with community validation and benchmarking guide improvements 3 people 2026-04-01 11:25:15 -05:00
• 6f152022e4 Restructure README for upstream-friendly framing 3 people 2026-03-31 19:17:40 -05:00
• 60cfc21cae Add MoE V-compression frontier paper, update recommendations 3 people 2026-03-31 13:46:36 -05:00
• 81360216b5 Fix wikitext download link in getting started guide 3 people 2026-03-31 13:38:15 -05:00
• 14280a6f91 Add missing branch checkout step to getting started guide 3 people 2026-03-31 13:07:45 -05:00
• 7d2cce3298 Add Vulkan validation to asymmetric K/V paper 3 people 2026-03-31 12:48:56 -05:00
• f89daef75c Fix addendum accuracy: dates, attributions, numbers 3 people 2026-03-31 11:49:10 -05:00
• 026da4bb1c Add independent validation addendums to all papers 3 people 2026-03-31 11:33:44 -05:00
• cecb66face Add paper: Asymmetric K/V Cache Compression 3 people 2026-03-31 11:28:11 -05:00
• c122fd0320 Add Sparse V and Boundary V sections to getting started guide 3 people 2026-03-31 11:13:55 -05:00
• cdbe5d5637 Add independently validated findings to README intro 3 people 2026-03-31 11:05:23 -05:00
• 79b42ab074 Make getting started and config links prominent at top of README 3 people 2026-03-31 11:03:36 -05:00
• c2a1bd8cb6 Add getting started guide link to README, fix iogpu value to 90% 3 people 2026-03-31 11:02:17 -05:00
• 7f5043ba2f Add getting started guide for new testers 3 people 2026-03-31 11:00:21 -05:00
• fdd5f6934d Update GPU memory cap recommendation to 90% (was 96%) 3 people 2026-03-31 10:36:15 -05:00
• 1e5eb03156 Fix context wall docs: only iogpu.wired_limit_mb is needed 3 people 2026-03-31 09:51:01 -05:00
• 88bf8ff092 Update stress test paper with 104B results and context wall fix 3 people 2026-03-31 09:47:27 -05:00
• 74f6e1bbba Add Command-R+ 104B stress test results, update recommendations 3 people 2026-03-31 08:06:54 -05:00
• 07e91a705b Remove accidentally staged files from prior commit 3 people 2026-03-30 23:25:17 -05:00
• ff28df2f60 Rename stress test paper, update references 3 people 2026-03-30 23:10:40 -05:00
• 0c6259bde4 Update recommendations with 70B, CUDA, AMD data + add stress test paper 3 people 2026-03-30 23:03:46 -05:00
• 715c90c1a2 Update architecture, project structure, roadmap, add Boundary V section 3 people 2026-03-30 22:01:42 -05:00
• 11a56625a2 Note AMD results are first attempt, no tuning 3 people 2026-03-30 21:18:20 -05:00
• 49889acb5f Add AMD RDNA 4 benchmarks and Windows HIP setup guide 3 people 2026-03-30 21:15:34 -05:00
• c819477692 Add block-size experiment paper and update docs 3 people 2026-03-30 13:33:17 -05:00
• 86bcbbe59a Add norm correction to Python prototype quantization path 3 people 2026-03-30 10:57:38 -05:00
• 13892d5e2c Refine Boundary V paper: scope to pure-attention models, document hybrid bug 3 people 2026-03-29 23:10:03 -05:00
• 90d1c5a297 Add boundary width vs precision addendum to layer-aware V paper 3 people 2026-03-29 21:16:18 -05:00
• 143303a35f Add M1 Max asymmetric K/V benchmarks and memory data from @mariotomich 3 people 2026-03-29 19:09:17 -05:00
• 1a941ac5d2 docs: Boundary V (experimental) — layer-aware V compression paper + recommendations 3 people 2026-03-29 17:39:58 -05:00
• 0935dcadc7 Add community hardware benchmarks: RTX 3090 CUDA + M1 Max Metal 3 people 2026-03-29 16:21:10 -05:00
• b50da4877b chore: consolidate NIAH logs into proof/niah/ 3 people 2026-03-29 15:02:52 -05:00
• 1d9df98f83 docs: fix asymmetric-kv-discovery accuracy 3 people 2026-03-29 14:43:02 -05:00
• de5f64ea77 docs: asymmetric K/V recommendations + investigation writeup 3 people 2026-03-29 14:36:45 -05:00
• 56a1604ff7 Experiment #25 result: context-adaptive dispatch negative on 1.5B 3 people 2026-03-29 00:00:05 -05:00
• e6ccc917e4 Experiment #23 result: specialized kernel = no improvement 3 people 2026-03-28 23:19:31 -05:00
• 61a13330ef docs: add turbo2 (2-bit, 6.4x compression) to README 3 people 2026-03-28 23:00:00 -05:00
• 9ecb478926 Add Phase 2 experiments #23-#25: kernel/format/pipeline level optimization 3 people 2026-03-28 22:37:52 -05:00
• c8858b84bb Correct M5 SMEM results: small regression, not catastrophic 3 people 2026-03-28 22:34:20 -05:00
• 824be6302e SMEM pre-dequant on M5 Max: -77 to -80% regression. Dead approach. 3 people 2026-03-28 22:21:57 -05:00
• bdb48d417e Correct SMEM pre-dequant results: +2-4% on small models, not -51% 3 people 2026-03-28 21:07:13 -05:00
• d38742e3f9 docs: add TODO section to cross-model validation for deferred tests 3 people 2026-03-28 20:54:00 -05:00
• 5c00b4eed1 Add approaches #20 + #22 results: 20 approaches tested, M2 ceiling confirmed 3 people 2026-03-28 20:43:03 -05:00
• fefd713cbf Add approaches 16h + 19 results: centroid LUT not the bottleneck 3 people 2026-03-28 20:23:11 -05:00
• af7bdf3d0c docs: complete cross-model rerun with ISWA fix + Gemma outlier log 3 people 2026-03-28 19:15:10 -05:00
• 5eade98036 docs: full ISWA investigation story in cross-model-validation.md 3 people 2026-03-28 18:57:26 -05:00