turboquant_plus

starred/turboquant_plus

0 Fork 0

mirror of https://github.com/TheTom/turboquant_plus.git synced 2026-10-22 02:53:13 -05:00

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

ba52ad107d README: mlx-swift-lm merged upstream (Apple PR #232) main TheTom 2026-07-20 15:08:51 -05:00
ace46fdb46 docs(rematch): turbo3/turbo2 now on the fused MMA decode path (bit-exact, flat at depth) docs/turbo4-rematch-jun2026 TheTom 2026-06-26 07:59:28 -05:00
b650b3e17a docs: turbo4 4.125 bpw + Jun-2026 rematch writeup; pin 3/4-bit centroid values TheTom 2026-06-26 07:30:43 -05:00
7f601a1357 README: drop RotorQuant from adopters table TheTom 2026-06-12 10:08:36 -05:00
3611a2d8c3 README: add credited community adopters to Run It Today table TheTom 2026-06-12 10:05:49 -05:00
31a46576a2 README: note upstream llama.cpp Hadamard KV rotation (21038 + FWHT kernels) TheTom 2026-06-12 09:54:07 -05:00
b77dbe4ecb docs: fix relative links in benchmarks.md after move TheTom 2026-06-12 09:42:58 -05:00
38aa2ae567 docs(readme): restructure around engine ecosystem table TheTom 2026-06-12 09:42:39 -05:00
7f04c8a112 docs: move status history to docs/changelog.md TheTom 2026-06-12 09:41:24 -05:00
6405f083af docs: move MLX framework port section to docs/mlx-port.md TheTom 2026-06-12 09:41:24 -05:00
1d54d2cd7c docs: move benchmark tables to docs/benchmarks.md TheTom 2026-06-12 09:41:14 -05:00
a6f4ef09f8 README: headline vLLM upstream TurboQuant merge (PR 38479) TheTom 2026-06-12 09:32:39 -05:00
7d69ed88e1 README: link lablup/mlxcel TurboQuant+ KV cache port TheTom 2026-06-12 08:27:19 -05:00
846590dda0 torquant-certificate: empirical eval of learned-rotor headroom on real K test/villines-probe-validation TheTom 2026-06-11 15:52:00 -05:00
9eb0341091 docs: add NOTICE file and per-file Apache-2.0 copyright headers TheTom 2026-06-11 10:08:15 -05:00
2f79598764 villines-probe addendum: bias hypothesis falsified, offset-ratio mechanism generalizes TheTom 2026-06-10 10:45:48 -05:00
3391e9f227 docs: link Atlas in the downstream-engines header TheTom 2026-06-10 09:18:50 -05:00
4dad409faa docs(villines-probe): T1-T4 reproduction + validation paper TheTom 2026-06-01 08:56:52 -05:00
280b466f7c docs(readme): mark QJL as reference-only, point to llama-cpp-turboquant for prod TheTomandClaude Opus 4.7 2026-05-28 11:05:51 -05:00
0cb20bca02 fix(qjl): use orthogonal projection and sqrt(d) scale factor (#93) Yash Raj Pandey 2026-05-28 11:56:04 -04:00
c46f6b9c43 docs(papers): add block-selector sparse attention WIP log
TheTom Claude Opus 4.7
2026-05-18 10:26:56 -05:00
8afc4bf3ea fix: K-side norm accounting — count both vector_norm AND residual_norm ship/pr-90-curated TheTomandClaude Opus 4.7 2026-05-09 10:46:30 -05:00
0ca5bcc303 ci: drop lint workflow, keep ruff config brosequist 2026-05-09 04:51:41 -04:00
3e37572a29 chore: add ruff linting to pyproject.toml and CI workflow brettandClaude Sonnet 4.6 2026-04-01 16:22:24 -04:00
f23570e0d9 test: add correctness and round-trip tests for fast rotation functions brettandClaude Sonnet 4.6 2026-04-01 16:22:10 -04:00
107462533b test: document QJL regression in test_turboquant_improves_over_polarquant brettandClaude Sonnet 4.6 2026-04-01 16:19:07 -04:00
b75813b0e1 fix: V-norm in memory_stats, SeedSequence PRNG, MSE compressed_size_bits brett 2026-05-09 10:42:53 -05:00
1224fef358 docs(papers): add longctx-1m-and-triattention + TriAttention V3 §10 addendum TheTomandClaude Opus 4.7 2026-05-08 13:21:49 -05:00
28e5178baa [paper] Remove broken script-dir reference from why-mse-fails TheTomandClaude Opus 4.7 2026-05-05 13:07:48 -05:00
9b22598e7f [paper] Clean references: drop /tmp scratch paths, point to in-repo scripts TheTomandClaude Opus 4.7 2026-05-05 12:06:14 -05:00
f71ec25f5c [paper] Add external confirmations: jagmarques (E8 VQ asymmetric) + quantumaikr (quant.cpp QJL-zero) TheTomandClaude Opus 4.7 2026-05-05 12:02:28 -05:00
0ca1439b3c [paper] When better reconstruction breaks models — MSE is not a valid proxy for KV cache quantization TheTomandClaude Opus 4.7 2026-05-05 12:00:49 -05:00
80fc2f4c84 [paper] Appendix A: document geometric-mean algebraic relation TheTomandClaude Opus 4.7 2026-05-04 16:31:51 -05:00
75ad729303 release(refract): v0.3.2.3 — MLX KLD numerical stability + walkthrough fixes v0.3.2.3 TheTomandClaude Opus 4.7 2026-05-03 20:17:24 -05:00
87b95b8e73 release(refract): v0.3.2.2 — --prompts defaults to bundled v0.3.2.2 TheTomandClaude Opus 4.7 2026-05-03 20:10:25 -05:00
b3252f6fe7 [paper] EDEN-S terminology fix per @amitport feedback in #89 v0.3.2.1 TheTomandClaude Opus 4.7 2026-05-03 15:29:08 -05:00
ff624e417b docs: link EDEN optimal-S response from README TheTomandClaude Opus 4.7 2026-05-03 12:19:37 -05:00
d271744554 [paper] Re-examining EDEN's Optimal-S Critique of TurboQuant TheTomandClaude Opus 4.7 2026-05-03 12:12:00 -05:00
8bf48f62c0 package(refract): rename to refract-llm, bump to v0.3.2.1, ship bundled assets TheTomandClaude Opus 4.7 2026-05-02 12:06:17 -05:00
64c19e76d5 fix(refract): null skipped axis score in JSON + 'skip' in compare table TheTomandClaude Opus 4.7 2026-05-02 10:47:19 -05:00
728cd8479b fix(refract): skipped axes don't inflate composite + report engine name TheTomandClaude Opus 4.7 2026-05-02 10:46:01 -05:00
8af12b360f test(refract): cover REFRACT_LLAMA_EXTRA_FLAGS escape hatch TheTomandClaude Opus 4.7 2026-05-02 09:59:33 -05:00
70a5bef212 feat(refract): REFRACT_LLAMA_EXTRA_FLAGS env var for constrained-VRAM users TheTomandClaude Opus 4.7 2026-05-02 09:56:00 -05:00
b331b8ca83 test(refract): refresh backend tests for production vLLM/SGLang status TheTomandClaude Opus 4.7 2026-05-02 09:13:57 -05:00
337f7c7696 docs(refract): QUICKSTART now leads with clone-and-install Step 0 TheTomandClaude Opus 4.7 2026-05-02 07:51:23 -05:00
eb0ebe3573 docs(refract): vLLM + SGLang backends production status across all docs TheTomandClaude Opus 4.7 2026-05-02 07:43:07 -05:00
6d5ec7de36 docs: rewrite cross-engine paper as facts-only TheTomandClaude Opus 4.7 2026-05-02 07:34:59 -05:00
d231c4c469 docs: cross-engine REFRACT bench paper on MI300X (Qwen3.6-35B-A3B) TheTomandClaude Opus 4.7 2026-05-02 07:19:14 -05:00
19ec3f4164 refract: batch PLAD ref/cand by KV config for memory-pressured backends TheTomandClaude Opus 4.7 2026-05-01 20:47:11 -05:00
4fb0c1c586 refract: dispatch tokenize_to_ids to active backend (R-NIAH unblock for MLX/vLLM/SGLang) TheTomandClaude Opus 4.7 2026-05-01 19:01:33 -05:00
08c8c86b35 refract: batch trajectory ref/cand by KV config to support evict-on-swap TheTomandClaude Opus 4.7 2026-05-01 15:56:32 -05:00
515822b5d6 refract: vLLM evict-on-key-change + MLX maybe_quantize import fix TheTomandClaude Opus 4.7 2026-05-01 14:54:49 -05:00
ba7f5e8e7f refract: implement vLLM and SGLang backends TheTomandClaude Opus 4.7 2026-05-01 11:34:49 -05:00
91ad2fe581 Lower the bar for REFRACT alpha: zero non-stdlib deps + cleaner pyproject TheTomandClaude Opus 4.7 2026-04-30 17:35:21 -05:00
4ec3dda5cf tests: MLX backend coverage via mocked mlx + mlx_lm TheTomandClaude Opus 4.7 2026-04-30 17:29:46 -05:00
d952a45745 selftest: detect shared-library load failures (Linux/Windows) TheTomandClaude Opus 4.7 2026-04-30 17:26:31 -05:00
c658eba453 REFRACT: 92% test coverage (297 new tests) + bug fixes + symturbo regen TheTomandClaude Opus 4.7 2026-04-30 17:06:22 -05:00
791fa86c8e README + QUICKSTART: explicit "how to generate the HTML report" section TheTomandClaude Opus 4.7 2026-04-30 16:04:44 -05:00
01c22235bc HTML report: full design refresh with dark-mode toggle TheTomandClaude Opus 4.7 2026-04-30 16:03:19 -05:00
b18a2d3cef R-NIAH: --rniah-up-to N knob for doubling step-up lengths TheTomandClaude Opus 4.7 2026-04-30 15:42:59 -05:00
43a4d2e9cf Auto-fetch wikitext-2-raw corpus + haystack TheTomandClaude Opus 4.7 2026-04-30 15:41:04 -05:00
a469d8e6ac HTML report: spell out acronyms, sanitize paths, synthesize repro for old JSONs TheTomandClaude Opus 4.7 2026-04-30 15:37:56 -05:00
c034e2f6af REFRACT v0.3.2 audit fixes from M2 mini fresh-install test TheTomandClaude Opus 4.7 2026-04-30 15:33:17 -05:00
a4c59a0b05 REFRACT v0.3.2: alpha release — 4 axes, 4 subcommands, 3 backends, 9 docs refract-v0.1 TheTomandClaude Opus 4.7 2026-04-30 15:13:04 -05:00
baebe5745f [paper] TriAttention V3: vLLM/AMD MI300X port addendum TheTomandClaude Opus 4.7 2026-04-30 12:20:12 -05:00
2d8a2472f7 REFRACT v0.1.3: defensive patch round (tests + fail-loud + honest docs) TheTomandClaude Opus 4.7 2026-04-29 20:36:25 -05:00
f2c3534654 REFRACT v0.1.2: subprocess utf-8 decode + true model-token GTM TheTom 2026-04-29 19:59:03 -05:00
a6b030e2ef REFRACT v0.1.1: fix four bugs that made the v0.1 matrix unusable TheTom 2026-04-29 19:22:18 -05:00
2e78787cb0 Promote refract to top-level package (refract/, not turboquant/refract/) TheTom 2026-04-29 18:41:02 -05:00
1197aa067e Add REFRACT v0.1: reference-anchored KV-quant quality oracle TheTomand 2026-04-29 18:35:34 -05:00
0d2c723079 Add attn-rotation + PPL artifact paper TheTom 2026-04-29 18:15:39 -05:00
3252b46900 Add DFlash self-draft investigation paper (31 experiments) TheTom 2026-04-25 12:55:58 -05:00
0398470f65 fix: point mlx-swift-lm links to alpha branch
TheTom Claude Opus 4.6
2026-04-23 12:34:41 -05:00
8d8ba1f7a6 docs: link vllm-swift in Apple Silicon section
TheTom Claude Opus 4.6
2026-04-23 12:33:18 -05:00
cf0caf6174 add prebuilt binaries section to README TheTom 2026-04-21 16:12:34 -05:00
c13be2e804 Add commercial support link to README nav
TheTom Claude Opus 4.6
2026-04-20 14:29:58 -05:00
422324e0c8 docs: incorporate Eric's feedback on turbo decode and Python parity
TheTom Claude Opus 4.6
2026-04-10 14:33:40 -05:00
fdb56e5775 docs: clean up pending items — PPL comparison done, update future work
TheTom Claude Opus 4.6
2026-04-10 14:25:58 -05:00
0cc2f20833 docs: add repo links to PR14 benchmark brief
TheTom Claude Opus 4.6
2026-04-10 14:21:33 -05:00
bc59e1499d docs: add MLX Swift PR14 Gemma 4 benchmark brief
TheTom Claude Opus 4.6
2026-04-10 14:18:43 -05:00
8a456b84e3 docs(triattention-v3): add Section 4.8 boundary skip negative result TheTomandClaude Opus 4.6 2026-04-09 18:03:36 -05:00
1e15bdbbed docs(triattention-v3): add Methodology section and renumber TheTomandClaude Opus 4.6 2026-04-09 15:45:29 -05:00
7fc9fafe94 docs: add TriAttention V3 paper TheTomandClaude Opus 4.6 2026-04-09 15:33:37 -05:00
ad28d36534 docs: add Apple Silicon callout pointing to ekryski/mlx-swift-lm
TheTom Claude Opus 4.6
2026-04-09 14:07:52 -05:00
7c716e0156 Remove nihilistau reference from asymmetric K/V paper
TheTom Claude Opus 4.6
2026-04-06 22:23:31 -05:00
aefb66ce86 docs: update MLX speed numbers — 97-100% baseline with delegated KVCache
TheTom Claude Opus 4.6
2026-04-06 15:25:54 -05:00
48984ee2e3 docs: add MM-NIAH benchmark results + mlx-vlm Quick Start TheTomandClaude Opus 4.6 2026-04-06 11:35:48 -05:00
aa8927c2b8 docs: compact mode uses C++ sdpa_vector_qv Metal kernel (precision bug fixed) TheTomandClaude Opus 4.6 2026-04-05 21:32:16 -05:00
e4fd48f99e docs: update MLX results with verified compact mode benchmarks TheTomandClaude Opus 4.6 2026-04-05 21:25:33 -05:00
97ab9c8d54 docs: fix MLX compact description — uses mx.quantized_matmul not C++ kernel TheTomandClaude Opus 4.6 2026-04-05 21:24:11 -05:00
0459d47906 docs: update MLX section with compact mode + C++ Metal kernel TheTomandClaude Opus 4.6 2026-04-05 21:22:38 -05:00
adce8c186f Revert "docs: add compact mode to MLX section" TheTom 2026-04-05 19:39:44 -05:00
83c03a8639 docs: add compact mode to MLX section TheTomandClaude Opus 4.6 2026-04-05 19:39:22 -05:00
ffb990d2a8 docs: fix MLX section — lazy compression, remove auto-compress claims TheTomandClaude Opus 4.6 2026-04-05 19:22:30 -05:00
e556e2105b docs: update MLX results with accurate K+V benchmark data TheTomandClaude Opus 4.6 2026-04-05 18:52:12 -05:00
3931b3ba97 revert: remove unrelated niah_test.py change TheTomandClaude Opus 4.6 2026-04-05 18:40:36 -05:00
92d4840cf0 docs: fix MLX section — K+V compression, accurate labels, memory table TheTomandClaude Opus 4.6 2026-04-05 18:39:49 -05:00
210d388fa9 docs: update for K+V compression and continuous compression TheTomandClaude Opus 4.6 2026-04-05 18:32:05 -05:00
a4b38535fe docs: add V cache memory savings table at scale TheTomandClaude Opus 4.6 2026-04-05 18:20:37 -05:00
31b710f4a7 docs: update MLX results with verified benchmarks TheTomandClaude Opus 4.6 2026-04-05 18:14:02 -05:00

1 2 3 4 5