malaiwah/Qwen3.8-27B-EXL3-K5K6-context

Qualify Qwen v5 body-only claims and reconcile legacy protocol summaries

Add pinned QFS size–KL plot, highlighted deep link and downloadable provenance

Add pinned QFS size–KL plot, highlighted deep link and downloadable provenance

Fix credits: RTX 5090 tester is malaiwah (self), GLM-5.2-EXL3-TR3-MTP78 is malaiwah's checkpoint; jcartu contributed the recipe/runtime that informed it

Add collection summary and fix credits (add Martin Vit, Luke Alonso, Brandon Music; correct jcartu attribution to MTP78 recipe)

Sync retrospective corrections and Final Frontier limitations

dedupe serving-profiles section (old unmarked copy removed; marked copy is canonical)

serving profiles: balanced+ANY_BITS n=3, 24/24 needles, baked image, ctx 249600

card: fidelity profile re-measured after the K5/ANY_BITS cure - prefill 1,966 -> 2,988 tok/s (+52%), decode 228.3/104.1, KLD 0.003405 (better), n=3 boots, gate 9/9

card: retract 'statistically level with official FP8' — the two KLDs come from different-sized suites and balanced's CI low end (0.005302) sits above FP8's point estimate (0.005294), so equivalence is not supportable

card: add measured serving-profile comparison + charts (measured on the -hydrated sibling, provenance noted)

charts: add fidelity-vs-quants.png

charts: add profiles-throughput.png

charts: add profiles-tradeoff.png

Address a peer review generated by our own served model: correct Three archival mirrors to Seven on four cards (all seven were already cited), and state both the build tree (21 files) and published tree (23 files) on the S16-V card with the +49,271 B accounted to README.md and DOCS-SHA256SUMS

Terminal-Bench 2.1: measured results, attribution and caveats

Correct the README row format in DOCS-SHA256SUMS so every row verifies with sha256sum -c

K6-parity: 0.001634 matches GGUF Q6_K, and the registered interval miss is published beside it

In-suite replay floor 5.83e-04, int8 embed overlay measured as an opt-in KV lever

V2 runner depth schedule measured: +38.0 % aggregate decode at C8, at the cost of 14.1 % KV tokens

Byte axis corrected: GGUF text-only file vs our multimodal payload; body-vs-body and deployed-multimodal now published

Head attribution measured: the output head is <=5.28% of divergence, so the cross-engine level gap is not decomposed

Scratch-arena overlay: measured +17,874 KV tokens (265,122 -> 282,996, +0.60 GiB) on the physical 5090, opt-in overlay not part of the qualified digest

KV-dtype sweep landed: fp8 measured default, per-token-head family measured, int4 capacity lever, nvfp4 refusal; V2-runner static-depth note

shard-0 comparator: add gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 row, paired results, third archival mirror; fix collection link

master chart: add the gittensor NVFP4 RTX5090 row; rebind DOCS-SHA256SUMS

Method hardening: publish the corpus text (corpus/text/, 941 documents + manifest) and correct the recoverability claims (audit gap G1); add the absolute-resolution paragraph beside every v5 table (gap G2); link receipts/startup-times.json beside the cold-start prose

Chat template: byte-identity to upstream, the three 400-class restrictions, prefix-reuse guidance

Link the upstream livelock issue #52520 and the boundary PR #47272

Utilisation is a per-card measurement: 0.955 is what qualified on GPU-506a575d, a second physical RTX 5090 needed 0.956, the margin is about 0.01 GiB, and a startup OOM should be answered by raising utilisation 0.001 at a time rather than dropping the window

Rebind docs checksums to the refreshed all-measurements chart

Redraw the all-measurements chart with the NVFP4 ten-shard row and the v3-vs-v5 ratio band

the sibling rebuild was scored: paired -3.755e-06 [-2.854e-05, +2.062e-05], brackets zero, so the untested-expectation wording is replaced by the measurement on all six cards

name the conversion-capable image, not an untracked path

converter determinism: a rebuild is a sibling, not the published bytes

context card: state R1's zero as a 95 % upper bound rather than an absence (0.90 per thousand builds, 1.42 per thousand gather-branch builds), per GdnGateAtConcurrency

GDN #51812 overlay recommendation on the three prefix-caching recipes, its measured absence on the context edition, and the 24 GB-class block (corrected gate clause) on the context card

Prefix caching: promote the four-module image to the release unit; enable the cache on the three 8,192-token recipes, decline it at the native 262,144 window with the measured mechanism, and correct the KV cost model to the affine per-request law

Cards: NVFP4 on v5, v3-vs-v5 scaling, capture determinism, external cross-citation, all-measurements chart, reproduce section, upstream fixes, 5090 lever sweep

Card consistency pass: resolve 16 audit findings (hardware-qualified framing, resident-vs-checkpoint bytes, whole-tree download convention, hardware-labelled throughput, figure alt text); refresh DOCS-SHA256SUMS

Bind docs checksum to normalized attributes

Hardware-qualified on a physical RTX 5090; GGUF comparison; corrected utilisation

Publish 10.48M-position KLD run, exact tail, paired MMLU-Pro matrix

Bind docs checksum to normalized attributes

Publish corrected qualification card and evidence

Chart: context edition reaches native 262,144 with int8 embeddings

Card: int4 draft embedding table is free (acceptance 56.7% vs 56.1%); MTP's cost at native context is not weights

Card: choice table reflects native 262,144

Card: native 262,144 context on a 32 GB card via an int8 embedding table (+0.000065 KLD), verified by 3/3 needle retrievals from 227,334-token prompts

Card: post-selection qualification on a source-disjoint suite - 42/42 paired wins over official FP8, ranking preserved