feat: EXL3 LoRA cartridge support for MSRT additive quantization - #1

Closed

malaiwah wants to merge 508 commits into

mainfrom

feat/exl3-lora-cartridge

This pull request is big! We're only showing the most recent 250 commits

Commits on Aug 12, 2026

design: VLLM_FQ_CAPACITY_UTILIZATION — gpu-max-util analog for tier headroom; L2 two-ledger mechanics behind one operator knob

malaiwah
and
claude
committed
127a069
m4-swap: T3 PASS (maps read as data under CUDA graph — atomic path live), T4 PASS x3 bitwise incl. rollback; apply window 0.061/0.368 ms (1/8 pairs, toy)

malaiwah
and
claude
committed
b82d765
loader-v2: trust + lazy-encode design/usage note — VLLM_FQ_SOURCES[_MODE], attestation trust filtering (signers/predicates, countersignatures), K-fallback ladder + EncodeQueue drain CLI, decision-l…

malaiwah
and
claude
committed
1b6b6af
0c: fix layer extraction in window publish (tr3 prefix collision)

malaiwah
and
claude
committed
87c7914
0c: robust window publisher (fresh temp staging per K, no stale-glob pollution) — replaces the shell version lost to a killed heredoc

malaiwah
and
claude
committed
0d47739
0c: window-2 ring — capture layers 11-18 from preserved boundary, then K2 (fast-load tier priority) + K5, auto-publish

malaiwah
and
claude
committed
8547df7
docs: 14-build-findings — what the build taught us (addendum to 00-13)

malaiwah
and
claude
committed
0d9fffb
docs 00-13: additive build notes so the spec stops contradicting the evidence

malaiwah
and
claude
committed
968ff1f
runs: README as an evidence index; research README points at runs/ + public artifacts

malaiwah
and
claude
committed
118cbed
m0-seed: HF card corrected against the evidence; 14 picks up the fq_verify proof

malaiwah
and
claude
committed
8a2e107
docs: normalize relative paths in the build notes; 11 picks up the expansion proof

malaiwah
and
claude
committed
c7e942f
0c: priming-report — window-1 community-quant extraction sealed

malaiwah
and
claude
committed
b0c65aa
fq_verify: prove the community-reassembly claim (byte-identity + numeric similarity)

malaiwah
and
claude
committed
923ad18
test_fq_verify: probe for fq_assemble --trust-signer instead of assuming it

malaiwah
and
claude
committed
cbb697d
m4-swap: T5 PASS (no torn state observable, 6/6 abort points) + T6 PASS (4 ranks bit-identical over 50 intervals); resolver-backed swap staging

malaiwah
and
claude
committed
3848253
card: prominent tools-repo + research links at the top; push the doc-coherence corrections live (tracked copy had been ahead of HF)

malaiwah
and
claude
committed
91556dc
fq tools: peer-review findings 1, 2, 4, 6 — record for b29c0621f

malaiwah
and
claude
committed
280ce6f
0c: K2-to-completion ring (operator priority) — layers 19-78 in 8-layer windows, capture->encode->publish->prune; K5 deferred to backfill

malaiwah
and
claude
committed
1714fd1
card: all four tiers in the title (K4 arrives via community priming); maturity split — assembly verified vs runtime experimental

malaiwah
and
claude
committed
5c98db2
fq tools: mirror fq_fetch / fq_release / fq_trust from the public repo

malaiwah
and
claude
committed
e4d5fb2
card: retire two stale known-issues (manifest last-writer-wins + K2/K5 mislabeling — both fixed), pin assembly examples to the signer (tool now fails closed), document recipes/

malaiwah
and
claude
committed
106d499
M1/M2: overhead A/B + dryrun evidence — collector costs 3.7-5.0% on the proxy (GATE <0.5% NOT met as measured; proxy amplifies fixed per-step cost ~150x, must re-measure on GLM-5.2 at cc8); M2 loop…

malaiwah
and
claude
committed
86c7170
0c: self-driving campaign supervisor — runs unattended for the life of the rental

malaiwah
and
claude
committed
f53d3f6
0c: supervisor single-owner lock + concurrent-encode guard

malaiwah
and
claude
committed
2bbd550
0c: campaign could not publish — supervisor ran the publisher without HF credentials, hanging on an interactive auth prompt (57min futex wait, zero uploads). Publisher now self-loads ~/.fq_env and …

malaiwah
and
claude
committed
ca5bcab
0c: prune each window's capture right after ITS publish (was: only after the last tier, so the whole K2 pass accumulated captures and ate 240GB). Re-capture for later tiers is deterministic, so not…

malaiwah
and
claude
committed
40f93e3
fq_assemble: --force can no longer eat a directory it did not create

claude
authored and
malaiwah
committed
ce09fda
fq_verify: pin the signer, and stop calling "unverified" a pass

claude
authored and
malaiwah
committed
35ad646
fq_prime spot-check: absent evidence is a failure, and JSONL means all lines

claude
authored and
malaiwah
committed
c29e306
m0-seed: HF card is the front door now — fix what was wrong on it, add LICENSE + NOTICE

malaiwah
and
claude
committed
a333c56
fq tools: mirror the atomic fq_release publish mode from the public repo

malaiwah
and
claude
committed
9fd871e
fq_fetch: authenticate the header the plan is built from

claude
authored and
malaiwah
committed
691e4a2
0c: defer capture prune while any encode is reading it

malaiwah
and
claude
committed
b2d56b0
card: name the fq-manifest MISMATCH that incremental publishing causes

malaiwah
and
claude
committed
ff799b0
fq_eps: commit the budget-conserving uniform baseline (was uncommitted)

malaiwah
and
claude
committed
fd89010
fq_repack: sign the segment header digest (sync from public) — 194 passed

malaiwah
and
claude
committed
0f8245c
0c: per-tier encode guard — different tiers may run concurrently on disjoint GPUs

malaiwah
and
claude
committed
f4579dc
0c: defer capture prune per-window, not globally

malaiwah
and
claude
committed
a119b88
0c: GPU reservation file so a loading serve is not raided

malaiwah
and
claude
committed
3671943
m5: swap-evidence recorder (metrics timeline + domain-shift workload)

malaiwah
and
claude
committed
56c3838
m5: GLM-5.2 TP4 serve script (off/dryrun/live modes)

malaiwah
and
claude
committed
dff36b9
m5: dependency-free SVG evidence charts + structural tests

malaiwah
and
claude
committed
3c51796
m5: evidence-campaign orchestrator

malaiwah
and
claude
committed
afdc6b7
health: rewrite sweep for the current pipeline + growth detection

malaiwah
and
claude
committed
7380fab
health: honest sweep labels (idle/exited/quiet vs STALLED)

malaiwah
and
claude
committed
b5c5d73
rebase: upstream overlap analysis for the PR duplicate-work check

malaiwah
and
claude
committed
a299070
m5: default serve window to 32k for honest eval numbers

malaiwah
and
claude
committed
0ac6eae
0c: never regenerate segments the remote already has

malaiwah
and
claude
committed
8a389f4
docs(rebase): map vLLM fork hierarchy and GG v20 r34 provenance

malaiwah
and
claude
committed
9c946f6
rebase: decided PR target, and the pre-filled-base trap

malaiwah
and
claude
committed
9bfd50f
health: track the newest m5 serve log, not a hardcoded tag

malaiwah
and
claude
committed
12fbf3d
m5: adversarial peer review of the evidence recorder + charts, and its fixes

malaiwah
and
claude
committed
8866774
m5 review: MIN-8 (sweep's hardcoded m5 tag) was fixed concurrently in 972beb96d

malaiwah
and
claude
committed
8e1810f
m5-serve: reference Coder quant map + MTP78 corpus loader + convergence plan

malaiwah
and
claude
committed
454fe2e
m5: assemble + verify pure-K3 and mixed-K3/K5 GLM-5.2 checkpoints from segments

malaiwah
and
claude
committed
6b00627
m5: fix four peer-review criticals before the evidence run

malaiwah
and
claude
committed
ba09cff
m5: M0 boot gate PASSES on the real GLM-5.2

malaiwah
and
claude
committed
dc0e45a
m5: convergence scorer — does runtime routing rediscover a human's quant?

malaiwah
and
claude
committed
89e35b3
m5: K5 mixed tier exceeds SM120 shared-memory limit (measured)

malaiwah
and
claude
committed
384ea5d
m5: scenario-1 boot policy (observe mode)

malaiwah
and
claude
committed
730e108
0c: pivot tier order to K2 -> K4 -> K5

malaiwah
and
claude
committed
b48409c
gg-env: deploy script — the serve runs the ROOTFS copy, not the source tree

malaiwah
and
claude
committed
c889d7a
m5-serve: spec the forced-retier admin API (design only)

malaiwah
and
claude
committed
26e50d0
fq/m5: answer "is the fungible-quant design topology neutral?"

malaiwah
and
claude
committed
135a1b4
m5: first convergence number on real GLM-5.2 (preliminary)

malaiwah
and
claude
committed
d3d8aee
m5: tensor-loader compatibility matrix for the fungible-quant integration

malaiwah
and
claude
committed
9bf45a5
m5: loader-compatibility — note the suite count drifts with concurrent agents

malaiwah
and
claude
committed
603f39a
m5-serve: report — a missing K bpw never crashes the serve

malaiwah
and
claude
committed
47c3926
fq/m5: bind real gate mass in the collector; deploy base_router.py

malaiwah
and
claude
committed
3aa5282
m5: convergence on the real MTP78 corpus — 0.3938, 1.48x chance, 59% of human

malaiwah
and
claude
committed
f1223db
m5: move raw routing dumps to HF, keep derived scores in git

malaiwah
and
claude
committed
4bd590d
m5: RETRACT the corpus convergence numbers — driver replayed truncated stubs

malaiwah
and
claude
committed
f2e288b
m5: demo-1 runner, per-expert K4 coverage discovery, honest scope

malaiwah
and
claude
committed
e1dd7ec
0c: disk guard must WAIT on its window, not advance past it

malaiwah
and
claude
committed
c46f8b4
0c: MoE layers are 3-77, not 3-78 (layer 78 is MTP)

malaiwah
and
claude
committed
b44a4f2
health: coverage denominator is 75 layers (3-77), not 76

malaiwah
and
claude
committed
415277a
m5-serve: admin API spec — "As implemented" section

malaiwah
and
claude
committed
f343cd0
m5: corrected corpus convergence — 0.4176, 1.57x chance, 62% of human

malaiwah
and
claude
committed
713e6c7
m5-serve/heatmap: design decision for the live MoE expert heatmap

malaiwah
and
claude
committed
f2ac7f5
FQ: activation-matrix endpoint spec (GET /fq/heatmap)

malaiwah
and
claude
committed
7679f39
0c: clear stale capture state when a window's data was pruned

malaiwah
and
claude
committed
0b2d8e5
FQ heatmap spec: reconcile the reset control with DESIGN.md

malaiwah
and
claude
committed
b88d7f4
0c: recover from PARTIAL captures left by preemption

malaiwah
and
claude
committed
d68323c
m5: 4-axis corpus replay driver

malaiwah
and
claude
committed
6c9a313
m5: adversarial review of the demo-1 glue — 5 defects, fixed with tests
malaiwah
committed
47d745b
deploy-fq.sh: deploy entrypoints/serve/__init__.py too

malaiwah
and
claude
committed
e70896c
m5/heatmap: 4-panel per-axis activation figure + pairwise overlap matrix

malaiwah
and
claude
committed
766770d
m5/heatmap: one permutation row per line in the JSON sidecar

malaiwah
and
claude
committed
a3d2b73
m5-serve/heatmap: operator heatmap page, self-contained, with maths + render tests

malaiwah
and
claude
committed
37c6425
m5-serve/heatmap: interop tests against the real endpoint's own encoder

malaiwah
and
claude
committed
fd3bb5e
m5: four-axis convergence — count beats mass, code axis is not special

malaiwah
and
claude
committed
13ab2b1
m5-serve/heatmap: adversarial review — three "pretty and wrong" fixes

malaiwah
and
claude
committed
5975078
m5: flagship 4-axis figure from REAL data + pairwise overlap matrix

malaiwah
and
claude
committed
7f49140
m5: GSM8K quality baseline — 89.2% on the assembled 3.0bpw checkpoint

malaiwah
and
claude
committed
50a7b63
0c: reclaim disk after encode, not only after publish

malaiwah
and
claude
committed
a5e0e91
m5: demo-1 progressive boot — segments + policy, no assembled checkpoint

malaiwah
and
claude
committed
8c0bca5
pr: upstream submission materials for fungible quant — not opened

malaiwah
and
claude
committed
4e7f46e
pr: pin the test number to the committed tree, inventory the renders

malaiwah
and
claude
committed
9ed57c8
m5-serve/heatmap: first real renders + headless browser toolchain

malaiwah
and
claude
committed
a8cad62
m5: progressive boot degrades instead of dying; ranged reads retry

malaiwah
and
claude
committed
208cca6
m5-serve: rewrite the HF model card for the FQ segments repo

malaiwah
and
claude
committed
62f5c1e
docs(fq): record upstream issues #282 and #283

malaiwah
and
claude
committed
03dcf22
m5: serve-demo1 deploys before booting — stale rootfs has cost 3 boots

malaiwah
and
claude
committed
92c4fdd
m5: refuse to boot onto occupied GPUs

malaiwah
and
claude
committed
255096b
m5: reap stale GPU holders by DEVICE, not by argv pattern

malaiwah
and
claude
committed
d148254
m5: enable hf_transfer for whole-segment prefetch

malaiwah
and
claude
committed
8b0b452
m5-serve: Xet high-performance transfer, engine timeout, prefetch depth

malaiwah
and
claude
committed
09b5f82
m5-serve: authenticate Hub downloads, report prefetch config at boot

malaiwah
and
claude
committed
22dab4c
m5-serve: document the progressive download path for the PR

malaiwah
and
claude
committed
91392d8
Commit completed pre-preemption work: HF card edits, KV addendum, adv-review

malaiwah
and
claude
committed
cf29bc8
m5-serve: record measured bulk-fetch throughput (~190x) and the defect trail

malaiwah
and
claude
committed
7ceef25
m5-serve: expose VLLM_FQ_KEEP_LAYERS and report it in the boot banner

malaiwah
and
claude
committed
9375e2c
m5-serve: record what a cold progressive boot actually costs

malaiwah
and
claude
committed
fabbb1f
gg-env: mirror fq_reload with the fq_converge_layers RPC

malaiwah
and
claude
committed
226be64
tools/fq_reload: fq_converge_layers RPC (and drop a duplicate copy)

malaiwah
and
claude
committed
f90c708
tools/fq_reload: fq_converge_layers must not imply an install it never does

malaiwah
and
claude
committed
22dd6f1
m5-serve: floor-triggered fragment-cache guard

malaiwah
and
claude
committed
8cc73bc
m5-serve: first complete progressive load, and the envelope that priced it

malaiwah
and
claude
committed
28cc444
0c-campaign: clean pause once K4 reaches 75/75

malaiwah
and
claude
committed
35e1368
m5-serve: battle-test plan, FQ_FAST loader-iteration mode, warm cache default

malaiwah
and
claude
committed
605fb32
health/sweep: watch the log the serve actually writes, probe the real port

malaiwah
and
claude
committed
44dee16
prune-fragment-cache: tiered floor, absolute cutoff, no silent no-op

malaiwah
and
claude
committed
4f5d3b8
health/sweep: count distinct layers on one rank, not matched lines

malaiwah
and
claude
committed
68fd23a
m5-serve: bt_metrics — make the hot-restart claim falsifiable

malaiwah
and
claude
committed
626d3d9
BT-1 PASSES: GLM-5.2 serves from Progressive Tensors segments

malaiwah
and
claude
committed
199587b
bt_metrics: count cache hits as their own segment origin

malaiwah
and
claude
committed
30426dd
BT-2 PASSES + retract the allocator-residue claim

malaiwah
and
claude
committed
c29decc
BT-6 FAILS: the loop decides swaps it cannot apply

malaiwah
and
claude
committed
b868012
preflight: enforce on calibrated projections, correct two bad constants

malaiwah
and
claude
committed
ce5ee5f
BT-1-AND-2: correct the stale 8.06 GiB overhead figure to the measured 7.47

malaiwah
and
claude
committed
b316193
m5-serve: verify_retier — never trust the swap API's own answer

malaiwah
and
claude
committed
166441d
verify_retier: pair the promotion with a demotion automatically

malaiwah
and
claude
committed
df21879
m5-serve: document the layer-78 asymmetry as one constraint, not four patches

malaiwah
and
claude
committed
55937bd
health/sweep: keep a progress signal on warm boots

malaiwah
and
claude
committed
6a51759
BT-1-AND-2: the zero-residue result is now confirmed 16 times over

malaiwah
and
claude
committed
40dc302
m5-serve: M4 swap verified on a live serve (evidence)

malaiwah
and
claude
committed
eb5afb6
growth_supported: design, and jaccard_trace to settle the guard question

malaiwah
and
claude
committed
887a962
growth: adopt A+B — one slot per layer, one operation in flight, both directions

malaiwah
and
claude
committed
1cd3fc6
heatmap: decode fq-heatmap/1 properly, and a finding that bounds the thesis

malaiwah
and
claude
committed
a8173d8
FINDING: GLM-5.2's routing is too flat for online top-K re-tiering

malaiwah
and
claude
committed
504cf06
FLAT: the domain change is invisible in the top-K churn

malaiwah
and
claude
committed
7476fc2
heatmaps: commit all 9 rendered samples for review

malaiwah
and
claude
committed
dd154e5
FLAT-1 preliminary: 15x the sample buys +0.014 Jaccard

malaiwah
and
claude
committed
7e42e09
SELECTION SIGNAL: routing frequency is right, and the encoder's data proves it

malaiwah
and
claude
committed
a967b9b
R10: the capture is documented, and the allocator supersedes my reconstruction

malaiwah
and
claude
committed
eb57339
ROUTING-FLATNESS: retraction banner — the instability half was a guard bug

malaiwah
and
claude
committed
6c690de
serve: FQ_TAG/FQ_DEVICES_ENV so two instances can run concurrently

malaiwah
and
claude
committed
65383c8
LANGUAGE MOVES THE EXPERT SET — 0.321 vs a 0.879 same-language control

malaiwah
and
claude
committed
5d22976
health/sweep: report BOTH instances, and whether apply actually installed

malaiwah
and
claude
committed
07e7c52
reap-devices: kill by DEVICE, because argv matching has now failed both ways

malaiwah
and
claude
committed
0e57191
heatmap format: benchmarked 7 encodings; the win was gzip, and it was free

malaiwah
and
claude
committed
815cde1
The guard fix is vindicated live: jaccard 0.61 -> 0.931 and climbing

malaiwah
and
claude
committed
6f43252
serve: calibrate VLLM_FQ_JACCARD_FLOOR to 0.80 from measurement

malaiwah
and
claude
committed
ec803f2
Scenario 2: flat-K3 policy, and the reserve arithmetic that bounds it

malaiwah
and
claude
committed
e909d90
serve: stop discarding an explicitly-passed CUDA_MODULE_LOADING

malaiwah
and
claude
committed
da132f8
BT-6 PASSES: 64 swaps INSTALLED into a live model, verified from state

malaiwah
and
claude
committed
ce8f19d
health/sweep: count installs by moment, not by rank log line

malaiwah
and
claude
committed
969c01b
SESSION-STATE: resumable snapshot before context compaction

malaiwah
and
claude
committed
8a35110
fq: tensor-level FQ feasibility — online per-slice re-tiering is not worth it

malaiwah
and
claude
committed
57d5981
fq: run artifacts through the 19:31 preemption

malaiwah
and
claude
committed
5f314ef
fq: the two-tier limit is a code constant, not a hardware one — measured

malaiwah
and
claude
committed
68eeab8
fq: cold-start handoff doc, and the concurrent-boot OOM that explains two deaths

malaiwah
and
claude
committed
03e2bc6
fq: HANDOFF — the complete open-work register, all 39 items

malaiwah
and
claude
committed
e3f3e19
BT-6c PASSES: the loop installs repeatedly, invalid swap = 0

malaiwah
and
claude
committed
e215331
health: surface cgroup memory and per-worker RSS in the sweep

malaiwah
and
claude
committed
3704d89
fq: handoff final for this session + BT-6c run artifacts

malaiwah
and
claude
committed
1bb2e6b
fq: final BT-6c artifacts at handover

malaiwah
and
claude
committed
2c58139
docs: add CHANGELOG.md summarizing all features and fixes on the branch

malaiwah
and
claude
committed
6ef0d79
poc: additive residual encoding for EXL3 trellis K2→K3→K4 — measured

malaiwah
and
claude
committed
28e81a5
poc: accurate additive residual measurements with real EXL3 trellis on RTX 5090

malaiwah
and
claude
committed
078e87e
poc: optimization opportunities — Lloyd-Max, cross-expert SVD, codebook comparison

malaiwah
and
claude
committed
05cf24e
chore: add .poc-venv to .gitignore

malaiwah
and
claude
committed
fe9208c
research: comprehensive PoC v4 — all 6 optimization ideas measured

malaiwah
and
claude
committed
ef04c46
research: v6/v7 — continuously variable per-expert sparse quantization on GLM-5.2

malaiwah
and
claude
committed
a4b827f
research: v8/v9 — BitsMoE decomposition + K4 tile K5 + full 3-tier (GLM-5.2)

malaiwah
and
claude
committed
1ad9ef8
research: v10 — AlphaQ PL_Alpha_Hill allocation (doesn't help for GLM-5.2)

malaiwah
and
claude
committed
8d6005e
research: comprehensive findings — tile-level 3-tier K3/K4/K5 is best method

malaiwah
and
claude
committed
f547a02
research: v11 — DP-optimal tile assignment + bitmap entropy analysis

malaiwah
and
claude
committed
038d6fd
research: v12 — proxy-based tier assignment (variance as fast proxy)

malaiwah
and
claude
committed
07d7091
research: round 2-4 literature review — RRQ, HyperQuant, ICQuant, Q-Palette, BitsMoE, AlphaQ, TileQ, MxMoE

malaiwah
and
claude
committed
e7d131f
research: final summary — tile-level 3-tier K3/K4/K5 is best fungible method

malaiwah
and
claude
committed
0ee87e6
research: v13 — cross-layer analysis (layers 10 & 40 are identical)

malaiwah
and
claude
committed
17f731b
research: round 5 literature — MXFP4, CodeQuant, MoBiQuant, GAMMA, WUSH

malaiwah
and
claude
committed
c28c6eb
research: v14 — tile difficulty is spatially random (no clustering)

malaiwah
and
claude
committed
0d1c8c6
research: round 5b — Leech lattice (LLVQ), entropy-constrained quantization, neural weight compression survey

malaiwah
and
claude
committed
8e32e49
research: v15 — complete smooth Pareto frontier (3.0-5.0 bpw, 0.1-bit steps)

malaiwah
and
claude
committed
2db2ca3
research: v16 — extended Pareto to 6.0 bpw with 4-tier K3/K4/K5/K6

malaiwah
and
claude
committed
ff04f3e
research: final Pareto frontier — continuously variable 3.0-6.0 bpw

malaiwah
and
claude
committed
8b839ef
research: complete report — 4-tier tile-level K3/K4/K5/K6, 3.0-6.0 bpw, 32 papers, 16 PoCs

malaiwah
and
claude
committed
6ef1dda
research: round 6 literature — GLVQ, MoBiQuant, expert merging, tensor networks, polar quantization

malaiwah
and
claude
committed
5550194
research: v17 — bitmap compression analysis

malaiwah
and
claude
committed
97e1596
research: v18 — progressive bitstream + quantized benefit (0% loss at 4-bit)

malaiwah
and
claude
committed
d6a1913
research: implementation guide — tile-level 4-tier fungible quantization

malaiwah
and
claude
committed
b598d5e
research: round 7 — embedded TCQ theory, FLUTE kernel, final status

malaiwah
and
claude
committed
81d72f1
research: v19 — per-tile rotation diversity (no effect)

malaiwah
and
claude
committed
09b9db2
research: round 8 — joint pruning+quantization, stochastic rounding, final thoughts

malaiwah
and
claude
committed
75062b7
research: round 9 — NVIDIA cuDNN Grouped GEMM+Quant (SM100+) hardware path

malaiwah
and
claude
committed
3ce574f
research: v20 — K5 residual decomposition comparison

malaiwah
and
claude
committed
d83348a
research: round 10 — TurboQuant theory + final conclusions (45+ papers, 20 PoCs)

malaiwah
and
claude
committed
efaa094
research: v21-v22b — corrected bpw labels + large Lloyd-Max codebooks

malaiwah
and
claude
committed
8361b46
research: v23b — per-tile Lloyd-Max codebooks (17-41% better than global)

malaiwah
and
claude
committed
10dd3a4
research: v24 — per-tile LM Pareto dominates global (1.6-70% improvement)

malaiwah
and
claude
committed
59517de
research: v25 — codebook clustering sweet spot at 64 clusters

malaiwah
and
claude
committed
f350e09
research: Fruit model analysis — per-expert heterogeneity is tier artifact

malaiwah
and
claude
committed
07252e1
research: v26 — universal normalized codebooks (cross-layer sharing = 1.000x)

malaiwah
and
claude
committed
14f4d81
research: definitive report — 26 PoCs, 50+ papers, universal codebook method

malaiwah
and
claude
committed
cf8aeca
research: v27b-v28 — K2 base tier + reshape bug fix + cross-layer re-verify

malaiwah
and
claude
committed
9c2b890
research: v29 — entropy coding saves 5-13% LM bits, PQ worse than scalar

malaiwah
and
claude
committed
2d566f9
research: v30 — c128 sweet spot, stacking always worse than single LM

malaiwah
and
claude
committed
4bb7815
research: round 11 + definitive report updated (v27-v30, 55+ papers, 30 PoCs)

malaiwah
and
claude
committed
4901128
research: v31 — definitive Pareto frontier (c128, 10 experts, gate+down)

malaiwah
and
claude
committed
e44cbcf
research: v32 — entropy-aware Pareto 14-49% better, BPDQ 4-23% worse

malaiwah
and
claude
committed
463185a
research: round 12 — BPDQ tested, final conclusions (60+ papers, 32 PoCs)

malaiwah
and
claude
committed
97e3d20
research: v33 — sparse/adaptive/tier-specific clusters all worse than uniform

malaiwah
and
claude
committed
c5af14b
research: v34 — trellis-on-residual 2-34× WORSE than Lloyd-Max

malaiwah
and
claude
committed
2cea6d9
research: round 13 — HARP, MSQ, QTIP, Proteus, NanoQuant, ReSpinQuant (65+ papers)

malaiwah
and
claude
committed
8a94c55
research: v35 — BREAKTHROUGH: rescaled trellis-on-residual 34-37% better than LM

malaiwah
and
claude
committed
8e574fb
research: v36 — definitive hybrid Pareto, K2+K4trsc 47% better at 6bpw

malaiwah
and
claude
committed
0e51d54
research: updated definitive report with v35-v36 rescaled trellis breakthrough

malaiwah
and
claude
committed
a533376
research: v37 — K2+K5trsc new best at 7bpw, K2+K2trsc matches K4 at 4bpw

malaiwah
and
claude
committed
f28410f
research: round 14 — successive refinement of TCQ, Drop-by-Drop, nested lattices

malaiwah
and
claude
committed
0ea1896
research: v38 — entropy-aware hybrid Pareto (LM entropy saves 0.56 bpw)

malaiwah
and
claude
committed
ec51d04
research: updated definitive report — v37-v38, K2+K5trsc, entropy-aware Pareto

malaiwah
and
claude
committed
8d90c55
research: v39 — THREE breakthroughs: 2-stage trellis 17% better, K6trsc beats LM at 8bpw

malaiwah
and
claude
committed
0c1d284
research: v40 — 3-stage rescaled trellis 2.3× better than LM at 8bpw!

malaiwah
and
claude
committed
4086aac
research: v41 — DEFINITIVE: MSRT beats LM at ALL bitrates 5-10 bpw

malaiwah
and
claude
committed
c6ab377
research: DEFINITIVE REPORT — MSRT beats LM at all bitrates (41 PoCs, 73+ papers)

malaiwah
and
claude
committed
4308d56
research: v42 — K2 base confirmed best, MSRT+LM hybrid 6× worse, universal

malaiwah
and
claude
committed
d5881eb
research: round 16 — RateQuant reverse waterfilling explains MSRT optimality

malaiwah
and
claude
committed
f5377b9
research: v43 — tile-level MSRT Pareto, 0-1.4% improvement from mixing

malaiwah
and
claude
committed
26e6ed2
research: v44 — MSRT 68-101× better than RRQ, cross-layer validated

malaiwah
and
claude
committed
70bbfec
research: round 17 — RRQ 68-101× worse than MSRT, GSQ/Proteus reviewed

malaiwah
and
claude
committed
afc961a
research: v45 — trellis indices are structured (16-bit), no entropy benefit

malaiwah
and
claude
committed
b5b36e4
research: round 18 — ECTCQ, embedded E-ECTCQ; trellis indices not entropy-compressible

malaiwah
and
claude
committed
787fb3a
research: v46 — systematic allocation search confirms K1-first pattern optimal

malaiwah
and
claude
committed
026a03a
research: round 19 — Drop-by-Drop, gradient-guided allocation; reverse waterfilling confirmed

malaiwah
and
claude
committed
3c6bc21
research: v47 — dithering 29-36% worse, per-tile K1 negligible

malaiwah
and
claude
committed
1cb9bcc
research: round 20 — dithering hurts TCQ, noise shaping N/A, interleaving not needed

malaiwah
and
claude
committed
a51e9bd
research: v48 — mul1 2.6× worse, up_proj universal, 70 experts CV=0.11%

malaiwah
and
claude
committed
4be4694
research: round 21 — LLVQ, BCJR-QAT, learned lattices; mcg codebook confirmed optimal

malaiwah
and
claude
committed
9923743
research: v49 — cross-layer validation + expert reordering (all ±0.03%)

malaiwah
and
claude
committed
ef2f865
research: MSRT additive cartridge feasibility + implementation plan

malaiwah
and
claude
committed
c69a80a
research: GLM-5.2 MSRT cost analysis — memory, performance, base K + additive bits

malaiwah
and
claude
committed
d5594f1
fix: correct cost analysis — MSRT advantage starts at 5bpw, not 3bpw

malaiwah
and
claude
committed
e630cda
research: MSRT cartridge via progressive-tensors + vLLM LoRA hot-swap

malaiwah
and
claude
committed
0d7bab7
research: v51 — MSRT cartridge matches native K4 within 3.1%

malaiwah
and
claude
committed
296b392
research: v52 — dual-cartridge MSRT (K2 base + tiered K1/K2/K3 cartridges)

malaiwah
and
claude
committed
f17cd8e
feat: EXL3 LoRA cartridge support for MSRT additive quantization

malaiwah
and
claude
committed
b1515dd
fix: shard_map must map projection names (gate_proj→w1, up_proj→w3, down_proj→w2)

malaiwah
and
claude
committed
d06c871
fix: extend EXL3 rank-sliced bitrate to accept K2

malaiwah
and
claude
committed
bb99f51