Part of the Qwen3.8-27B mixed-precision EXL3 (measured) collection.

Four EXL3/Trellis quants of Qwen/Qwen3.8-27B, each optimizing a different constraint. Every quality claim is measured against a BF16 teacher via KL-divergence — no unverified accuracy assertions.

Model	Download	Resident	v5 mean KLD	Context (32 GB)	TL;DR
K5K6-hydrated	21.61 GB	20.31 GiB	0.002760	~180k	Best fidelity, smallest download, fastest cold start
K5K6	30.60 GB	20.32 GiB	0.003210	~180k	Attention width is a launch-time knob (K6/K5/K4)
K5K6-context	20.70 GB	18.41 GiB	0.003509	262,144	Native 262k context, int8 embeddings, hardware-qualified on RTX 5090
K4	28.31 GB	17.89 GiB	0.010604	262,144	Smallest footprint, native context, no overlay needed

All four are VLM (image-text-to-text) and require the Gilded Gnosis vLLM runtime — they do not load in stock vLLM, SGLang, or transformers.

Qwen3.8-27B-K4 — EXL3 mixed-precision, the capacity edition

Superseded by malaiwah/Qwen3.8-27B-EXL3-K5K6

The successor spends the remaining memory budget on the MLP (gate/up K5, down K6) and measures 0.003210 mean KLD on the v5 held-out suite against this checkpoint's 0.010604 (0.007945 versus 0.029679 on the older overlap-corrected v3 subset), at 21.82 GB resident — 39 % below official FP8 on v5, 38 % below it on the v3 subset. It also has a quantized MTP draft head that works with speculative decoding (58.2 % acceptance, +101 % single-stream throughput). Prefer it unless you specifically need the smaller 19.21 GB footprint.

Two corrections to this card, from an independent review: the first published comparison used evaluation prompts drawn from the quantizer's own calibration corpus (re-measured held-out numbers are below), and reasoning_effort accepts only xhigh/medium/low. This checkpoint's lm_head also carries the mul1 codebook rather than the mcg implied by the documented -cb mcg, an artefact of that run's crash-and-resume; the successor's command produces mcg throughout. That is a claim about the recipe, not about the bytes — see Reproducing this quant.

Dense EXL3 quant of Qwen/Qwen3.8-27B built on one principle: spend 4 bits only where the two independent NVFP4 recipes for this architecture also spend 4 bits, and protect everything they protect — but protect it with Trellis, not with FP8.

• MLP (gate_proj/up_proj/down_proj, all 64 layers) → EXL3 K4, serialized, calibrated LDLQ.
• Attention (linear_attn.{in_proj_qkv,in_proj_z,out_proj} on the 48 linear layers, self_attn.{q,k,v,o}_proj on the 16 full-attention layers) → BF16 on disk, encoded to K6 at load time by the Gilded Gnosis vLLM fork's ONLINE_QUANT=exl3-b6 overlay, cached content-addressed on disk.
• lm_head → EXL3 K6 (6 bpw, serialized).
• embed_tokens, vision tower (27 blocks), MTP draft head, norms → BF16, untouched.

Measured resident weights: 17.89 GiB (19.21 GB) on a single RTX PRO 6000 Blackwell. The one NVFP4 build of this architecture whose resident weights were measured here under identical flags is unsloth/Qwen3.8-27B-NVFP4 at 21.34 GiB (22.91 GB), so this build holds 3.70 GB less resident weight (docs/22-results-iteration-2.md). Those are load-time device allocations. They are not the same quantity as the 21.92 GB of nvidia/Qwen3.6-27B-NVFP4 and the 23.42 GB of unsloth/Qwen3.8-27B-NVFP4 quoted elsewhere: those two are checkpoint sizes — serialized bytes on disk, never resident memory — and no resident figure has ever been measured here for the nvidia checkpoint, which is a different model generation (Qwen3.6). Per-role bit widths are not directly comparable across Trellis, NVFP4 and FP8, so the comparison that matters is the measured fidelity below, not a per-role precision claim.

Headline fidelity now comes from the v5 held-out suite — 5,120 contexts x 2,047 positions = 10,480,640 scored positions. This build measures 0.010604 mean KLD [0.009640, 0.011746] with 95.76 % top-1, against 0.005294 for Qwen/Qwen3.8-27B-FP8: on that suite official FP8 is the more faithful checkpoint, by +0.005310 [+0.004710, +0.006019] paired, with this build ahead in only 7 of 5,120 contexts. The direction is unchanged from v3 (0.029679 here versus 0.012798 for FP8 on the corrected subset); only the absolute values moved, because the v5 corpus mix is different. Every earlier receipt is kept below. With CUDA graphs enabled this build still serves faster than the NVFP4 checkpoint (55.39 vs 49.09 tok/s at C1). Vision works. The v3 artifacts are published as a dataset; the v5 run ships as receipts (receipts/kld5-*.json).

Position in the family: this K4 build has the highest divergence of the four published builds (0.010604 on the v5 suite, 0.029679 on the corrected v3 subset) and the smallest resident footprint (17.89 GiB). It reaches native 262,144 on a real 32 GB card with no overlay at all — 289,577 KV tokens measured on an RTX 5090 with MTP-3 and vision enabled. The context edition is now hardware-qualified on a physical RTX 5090 as well, at 265,122 KV tokens for 262,144 with MTP-3 and the full 8.4 MP image ceiling, measured at --gpu-memory-utilization 0.955 with all seven gates passing (receipts/qualification-5090-context.json), so between the two the trade is footprint and overlay-free simplicity against fidelity, not capability.

That 0.955 is a per-card measurement, not a constant. It is the value that qualified on one board, GPU-506a575d (32,607 MiB, 458 MiB of it held by the driver); a second physical RTX 5090 needed 0.956, missing at 0.955 by about 0.01 GiB (receipts/second-5090-datapoint.json). Two nominally identical boards differ in exactly two quantities no configuration can move — the driver's framebuffer reserve and the CUDA context size — and a 68 MiB perturbation in either was measured to be enough to flip a gate (receipts/qualification-24gib-capped.json → residual_risk_versus_a_physical_board), while one thousandth of utilisation is only about 32 MiB. So if a card refuses to start or OOMs at startup, raise utilisation by 0.001 at a time rather than dropping the window — and do not lower max_pixels to make room, because at fixed utilisation that enlarges the KV pool and makes the large-image case fail sooner.

Which of the four builds

Same architecture and tokenizer. The first KLD column is the v5 held-out suite (10,480,640 scored positions); the second is the older overlap-corrected 127-context v3 subset, kept for continuity. Absolute KLD is suite-specific, so the two columns are not comparable with each other — only the ordering within a column is. Capacity uses each card's documented profile: hydrated, online and K4 are real RTX 5090 MTP-3 tests; context is MTP-3 with an 8.4 MP cap, qualified on a physical RTX 5090 at utilisation 0.955. These profiles are not interchangeable (collection).

Byte and memory conventions for this table. The download column is whole-tree bytes — every published file of the artifact as its release evidence counted it (receipts/collection-index.json, serialized_bytes.whole_tree_bytes: hydrated 21,610,933,884 B, K5/K6 30,597,231,933 B, context 20,696,053,306 B) — and they are serialized bytes on disk, never resident memory. *This build's release evidence records no tree count, so its own row is the sum of its safetensors shards, 28,313,841,196 B, read from the published repository. The context edition's resident weight is measured twice: 18.41 GiB as run on the rental RTX PRO 6000 engine-budget proof and 18.19 GiB on the physical RTX 5090 at the qualified 0.955 profile. This table prints the larger figure deliberately, because receipts/vram-class-verdict.json elects 18.41 GiB for every class prediction; the 0.22 GiB gap is the rental-versus-5090 delta, not a change in the checkpoint.

Official Qwen/Qwen3.8-27B-FP8 is 28.51 GiB resident at 0.005294 on the v5 suite and 0.012798 on the corrected v3 subset, and runs on stock vLLM, which none of these do. On both suites it is more faithful than this K4 build and less faithful than the other three.

Why this shape

Trellis at K4 needs no per-group scale tensor, so 4-bit MLP costs 4.004 bpw here against NVFP4's 4.50 bpw. That saved 1.07 GB, plus the 1.80 GB from serving attention at K6 instead of FP8, is what pays for a lower footprint at higher precision. The checkpoint is larger than the NVFP4 ones because attention ships BF16 so the runtime can re-encode it — see Tradeoffs.

Distribution fidelity — v5 held-out suite, 10,480,640 scored positions

This is the headline fidelity evidence for the family, and it is the one measurement where this build loses to official FP8. Suite receipts/kld5-suite-manifest.json (schema qwen38-distribution-fidelity/6, suite token sha256 510541f6861b589d44932db253ec25d96d6daaeeee4ea2ab9b65329209482b88): 5,120 contexts x 2,047 positions = 10,480,640 scored positions over 842 source clusters, from a corpus of 941 documents / 70,348,971 bytes fetched by tools/fetch_corpus_v5.py (fetch log). All 941 documents were scanned at every position for exact normalized 12-token overlap with exllamav3 calibration data before selection, and the 44 that hit (43 code, 1 encyclopedic) were excluded whole, leaving 897 eligible documents — so contamination hits are 0 by construction. The suite is token-disjoint from the v4 suite (0/160 prior context hashes reachable), and its windows are exact-advance and non-overlapping — independently verified at 5,120/5,120 unique token hashes and 0 overlapping windows. Every candidate is scored body-only: both operands go through one shared BF16 LM head.

How closely these absolute numbers may be read. Each mean is a body-only replay value: both operands are projected through the one shared BF16 head, and the replay path is not the engine's own logit path. Replaying the unquantized model against its own live logits measures KL(live ‖ replayed) = 5.83e-04 — 32 v5 shard-0 contexts, 65,504 scored positions, context-bootstrap 95 % CI [5.15e-04, 6.64e-04], top-1 99.10 %, on the same suite, reference capture and shared BF16 head as the means above (receipts/replay-live-floor-v5.json), superseding the six-context v3 derivation of 6.54e-04 (receipts/v3-qualification-bf16.json), which its interval contains — and moving hidden-state storage from BF16 to fp32 moves a candidate's KLD by 5.6 % (docs/24). Absolute values are therefore within-suite numbers: they carry a ~6e-4 implementation offset plus a ~5 % storage systematic, and absolute differences below about 1e-3 are not resolvable. Both offsets are common-mode — every candidate replays through the identical path — so paired differences and orderings are the resolvable quantity: hydrated − online K5/K6 is −4.50e-04 [−4.69e-04, −4.33e-04] on 4,922 of 5,120 contexts (receipts/kld5-10M-paired.json), smaller than the replay floor and resolved because the floor cancels in the pairing. The floor is now derived inside the suite rather than on six out-of-suite v3 contexts, and the rule it licenses is unchanged. What it does not license: it is not a claim that candidate KLDs are 11 % smaller, and it does not let any single absolute mean be read more finely — the 5.83e-04 figure is a mean over 32 contexts whose own means span 3.09e-04 to 1.63e-03 with a worst single position of 0.2534. It is also not the cross-engine floor (0.000507), which is a different control. Method of record: docs/42-kld-method.md.

Paired per-context differences (source-cluster bootstrap, 10,000 resamples, seed 1, 842 clusters, receipts/kld5-10M-paired.json):

Stated plainly: official Qwen/Qwen3.8-27B-FP8 is markedly more faithful than this build, and it wins 5,113 of 5,120 contexts. That is the same direction the v3 receipts reported; ten million positions did not overturn it, they narrowed the interval around it. If you want distribution fidelity, take one of the other three builds or FP8; take this one for the 17.89 GiB footprint and native context without an overlay.

The widest single view of the evidence: A is the only panel where every family appears together (v5 shard 0, 512 contexts, 1,048,064 positions, ten candidates), B is the same suite's 1M → 10M ladder (five vLLM builds, 5,120 contexts at 10M), C1/C2 are the superseded corrected v3 (127 contexts, 259,969 positions) and the source-disjoint v4 (36 contexts, 73,692 positions) on separate y-axes behind a barrier, and D is turboderp's own published protocol (OpenWebText, 65,536 positions), which we have never run. Two rules travel with the figure: the cross-engine floor belongs to the llama.cpp rows only, and no ratio across panels means anything. Generated by tools/make_master_kld_chart.py, which reads every one of our values from receipts/ at runtime.

These numbers are re-derivable, not merely re-runnable. 5,240,320 scored positions — five candidates x 512 contexts x 2,047 positions — reproduce bit-for-bit across independent runs with separate model loads and different harness generations: every measured field identical, including the complete per-context arrays and the whole bootstrap block, with only the capture directory paths and the additively-added tail histogram differing (receipts/capture-determinism.json), and a third harness generation's unwindowed --score-from 0 control returns the same shard-0 means to the last digit (receipts/scored-window-offset.json). The scope is part of the claim and travels with it: one GPU, one driver, one pinned rootfs, enforce_eager=True, max_num_seqs=1, one context per forward, 512 MiB bf16 KV. It is not a claim that vLLM is bitwise deterministic in general — nothing here covers CUDA graphs, max_num_seqs > 1, chunked prefill with more than one chunk per context, other GPUs or drivers, or anything downstream of the logits.

The corpus mix moves absolute values for every candidate, so v5 numbers are not comparable to v3 numbers. This build reads 0.029679 on the corrected v3 subset and 0.010604 here; FP8 reads 0.012798 there and 0.005294 here. Nothing about the checkpoints changed between those two runs — the text did. Only within-suite ordering and the paired differences transfer across suites.

By stratum, this build's mean KLD is 0.007389 scientific, 0.008897 encyclopedic, 0.010356 multilingual, 0.011612 code and 0.014766 literary, over 1,024 contexts (2,096,128 positions) each. The cumulative mean is stable along the ladder: for the hydrated candidate the 1M / 2M / 5M / 10M checkpoints read 0.002700 / 0.002759 / 0.002699 / 0.002760, so the 10M figures are not a moving target.

Per-candidate receipts are receipts/kld5-10M-{hyd,k5k6,ctx,fp8,k4}.json (schema qwen38-kld-ladder-cumulative/2), built by tools/kld_aggregate.py from ten verified per-shard reports produced by tools/kld_ladder.sh: capture six models over 512 contexts, replay five candidates, verify, delete 64 GB of hidden states, move to the next shard.

Distribution tail. A mean and a top-1 rate can hide a tail, so here is the tail itself, measured on shard 0 of the same suite — 512 contexts, 1,048,064 scored positions — the identical contexts for all five candidates. Receipts receipts/kld5-1M-tail-{hyd,k5k6,ctx,fp8,k4}.json (schema qwen38-kld-ladder-cumulative/2, built by tools/kld_aggregate.py); this build's row is receipts/kld5-1M-tail-k4.json. Every qwen38-fidelity-report/2 replay accumulates a 560-bin log-spaced histogram of per-position KLD (KLD_HIST_LOG10_LOW=-12.0, KLD_HIST_LOG10_HIGH=2.0, KLD_HIST_BINS_PER_DECADE=40 in tools/fidelity.py) whose bin counts add across shards, which is what makes cumulative quantiles possible at all.

Stated plainly: this build's tail is heavier than official FP8's at every measured quantile. p99 is 0.1194 against FP8's 0.0531, p99.9 is 0.5555 against 0.2438, p99.99 is 1.870 against 0.812 — roughly 2.3x at each depth — and the exact worst position on this shard is 7.565 against 5.296. The exceedance counts say the same thing without any binning: 1.2604 % of positions exceed 0.1 against FP8's 0.3912 % (3.2x), and 0.03807 % exceed 1.0 against FP8's 0.00592 % (6.4x, 399 positions against 62). The tail ordering at p50, p95, p99, p99.9 and p99.99 is the same as the ordering of the means, so the mean was not flattering this build: it is last on the mean and last at every quantile, and all three K5/K6 builds are below FP8 throughout. Take this build for the 17.89 GiB footprint and native context, not for distribution fidelity — including at the tail.

Three limitations, stated rather than hidden:

• Not comparable across suites. Absolute KLD is suite-specific; see above.
• Cumulative percentiles come from one shard, not from all ten. The ten shard reports of the 10M run carry no token-level KLD histogram, so nothing could be recombined across them. The tail table above closes that gap on shard 0 (receipts/kld5-1M-tail-*.json), with bin-bounded quantiles — each receipt carries lower / upper / estimate, relative bin width about 5.6 % — and exact maxima and exceedance counts. Across all 10,480,640 positions only the means, the intervals, the paired results and one exact global maximum per candidate exist.
• Captures: the reference survived, the candidates did not. The five candidates' hidden states and the BF16 references for shards 1-9 were deleted shard by shard to fit 135 GB of scratch. The shard-0 BF16 reference was kept and is published, together with the suite, all ten shard views and 79 per-shard reports, so a new candidate can be scored against the identical contexts without recapturing the reference — see Reproduce this.

Against GGUF, measured on our suite

The comparator set used to stop at official FP8, which is a throughput format whose quality is Q4-to-Q5 class, so llama.cpp's Q8_0 and Q6_K are the honest bar. They have now been measured on our own suite, and for this build the result is unambiguous: every GGUF measured here beats it, including the smallest of the three. The one row it does beat is the checkpoint its readers actually weigh it against — unsloth/Qwen3.8-27B-NVFP4, the other 4-bit-weight-class build for this architecture, which has now been measured on the identical shard and is 2.9x this build's divergence.

Three GGUFs from unsloth/Qwen3.8-27B-GGUF@f1bfb127c64f7072bdd2cad55f258b9c8b2910fe were captured under llama.cpp pinned at commit ece963f41b0b02d7a0d61436ae365762c073a4c8 with tools/gguf_capture.cpp, which reads the post-final-norm state — the same mathematical point the vLLM hook takes, with bf16 rounding verified bit-identical to torch on 2,012,449 probe values — and scored against the same BF16 teacher through the same shared BF16 head, on shard 0 of the v5 suite: the same 512 contexts and the same 1,048,064 scored positions every row below saw. Manifests come from tools/gguf_manifest.py and each one carries the GGUF blob digest and the llama.cpp identity; the build script is tools/build_llamacpp.sh. Receipt receipts/cross-engine-comparator.json, per-candidate reports receipts/gguf-report-{q8_0,q6_k,q5_k_xl}.json.

The engine control, measured and not assumed. The unquantized BF16 GGUF captured in llama.cpp against the vLLM BF16 reference on identical token ids, the shared head and the same 512 contexts measures 0.000507 mean, 99.07 % top-1 and p99.9 0.0113 (receipts/gguf-report-engine-floor.json). It proves the engine is a confounder. KL is neither additive nor a metric, so the control cannot be subtracted and supplies no quantization-only upper or lower bound. Cross-engine rows compare complete pipelines. Four cells read —, for two different reasons. Two of them are the builds that ship BF16 attention for the runtime to encode at load, including this one (28.31 GB download, 17.89 GiB resident), so their disk bytes are not a like-for-like payload against a GGUF file; the payload figures are immutable_payload_bytes from receipts/collection-index.json (hydrated 21,610,916,123 B = 20.127 GiB, context edition 20,696,033,532 B = 19.275 GiB; the table truncates both to two decimals) and are serialized bytes, never VRAM. The others are the NVFP4 builds, for which we publish no serialized-byte receipt of our own: unsloth's 22.91 GB is measured resident weights and its 23.42 GB is a checkpoint size, gittensor's 18.77 GiB is likewise measured resident weights, and none of these is the payload quantity this column holds, so they get no x rather than an invented one. The FP8 figure is resident weights and is labelled as such.

The p99.9 column, and why it differs from the tail table above. These p99.9 values are each report's exact shard-0 p99.9 as the comparator receipt read them; the tail table above quotes the bin-bounded cumulative estimate from the 560-bin histogram, whose bins are about 5.6 % wide — this build reads 0.5555 there and 0.5576 here, hydrated 0.1319 and 0.1313, and each exact value lies inside the bin the estimate names. The two differ by construction, not by measurement.

NVFP4 on the identical shard — the comparison this build's readers actually make. unsloth/Qwen3.8-27B-NVFP4 at revision 9c73e2da is the other 4-bit-weight-class checkpoint for this architecture, and it is served by the same vLLM build as this row, so unlike the GGUF rows it carries no cross-engine term at all and is directly comparable to ours with nothing subtracted or estimated. On the same 512 contexts, the same 1,048,064 positions and through the same shared BF16 head, it measures 0.030115 mean KLD, 95 % CI [0.027637, 0.032965], 93.16 % top-1, median 0.009584, p95 0.10051, p99 0.33546, p99.9 1.6228, exact worst position 10.6285 and mean JSD 0.010104 bits (receipts/kld5-1M-nvfp4.json, with the run's own account in receipts/nvfp4-v5-measurement.json). That is 2.9x this build's 0.010345 at the same 4-bit weight class, 5.8x official FP8, 8.8x the context edition, 11.2x the hydrated build and 27.7x Q8_0 as measured; its p99.9 is 2.9x this build's 0.5576. Its own histogram brackets that quantile at [1.5849, 1.6788] with a 1.6128 point estimate, which contains the exact 1.6228 — the same construction difference described just above, in the one row where both numbers are published.

Paired per context, which is a stronger statement than any ratio of means: NVFP4 loses every one of 512 contexts, against both comparators it was paired against. +0.026706 against the context edition (95 % CI [+0.024465, +0.029285], 0 wins to 512) and +0.024918 against official FP8 ([+0.022756, +0.027424], 0 wins to 512) — not one context anywhere in the shard where it is the better of the pair (receipts/kld5-1M-paired-nvfp4.json). It was not paired against this build, so the 2.9x above stays a ratio of means and is not presented as a win count.

gittensor's "RTX5090" NVFP4, measured on the same shard because its card claims the 32 GB / 262,144-token axis by name. gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 at revision 69274a0d (ModelOpt W4A4 body, MTP and vision left BF16, FP8 KV cache baked into its config) is served by the same vLLM build as our rows, so it carries no cross-engine term. It measures 0.062163 mean KLD, 95 % CI [0.058491, 0.066360], 89.85 % top-1, p99.9 2.5911 — the weakest row on this table, at 2.1x unsloth's NVFP4 and 6.0x this build. Paired per context it loses every one of 512 contexts to this build (+0.051818 in this build's favour, 95 % CI [+0.048909, +0.055160], 512 wins to 0), every one of 512 to official FP8, and 511 of 512 to unsloth's NVFP4 (+0.032048, [+0.030711, +0.033583]) — the same weight format at 2.57 GiB less measured resident weight (18.77 vs 21.34 GiB, identical flags, engine-reported), which prices that memory saving honestly: roughly double the KLD. A bf16-KV control capture moves its mean by only +0.000365 [+0.000058, +0.000679], so its baked FP8 KV cache explains about 1 % of the gap to unsloth — the rest is the weight conversion itself. Its card's serving numbers (18.8 GB weights in VRAM, 275,941-token FP8 KV pool, 80.6 tok/s decode, native 262,144 on one 5090) are its own claims, which we did not run; our measured 18.77 GiB resident is consistent with the first of them, and consistency is not verification. Its only published fidelity evidence is a 20-item smoke that its own card says not to treat as scores (receipts/kld5-1M-gt5090.json, receipts/kld5-1M-paired-gt5090.json, full account with the checkpoint's composition, digests and mirror in receipts/gittensor-nvfp4-rtx5090.json; archival mirror malaiwah/Qwen3.8-27B-NVFP4-RTX5090-archival-69274a0d).

Where this build sits, stated plainly. Third from last, and last among the rows a reader choosing on fidelity would shortlist. Its 0.010345 mean is 2.6x UD-Q5_K_XL's net 0.003936, 6.8x Q6_K's net 0.001528, 17.9x Q8_0's net 0.000579 and 2.0x official FP8's 0.005197; its top-1 is the lowest of those rows at 95.91 %, and its p99.9 of 0.5576 is 2.6x UD-Q5_K_XL's 0.2144, 7.0x Q6_K's 0.0794 and 2.3x FP8's 0.2440. It is beaten by every GGUF measured here, including UD-Q5_K_XL, the smallest of the three at 18.83 GiB of serialized weight. What it is not beaten by is either 4-bit alternative: unsloth's NVFP4 measures 2.9x this build's divergence and gittensor's 6.0x, on the identical positions, in the same engine, with no floor to argue about. This build's argument was never fidelity — it is the 17.89 GiB footprint and native 262,144 context with no overlay — but that is a capacity argument, and a reader choosing on distribution fidelity should take a GGUF or one of the K5/K6 builds instead.

The two complete-pipeline observations for the family:

1. At the nominal 6-bit point, llama.cpp Q6_K measures 0.002035 and the vLLM hydrated build 0.002700.
2. At the nominal 5-bit point, the vLLM context edition measures 0.003409 and llama.cpp UD-Q5_K_XL 0.004444.

Neither comparison isolates format because the candidate engines differ.

Update 2026-08-16 — a near-equal-file-byte K6 build. The hydrated recipe with gate_proj and up_proj promoted K5 → K6 measures 0.001634 mean KLD [0.001541, 0.001742], top-1 98.248 % (k6-parity-kld.json). It beats hydrated by 0.001066 [0.000974, 0.001173] on 511 of 512 contexts for +1.348 GiB, and its complete vLLM pipeline measures 0.000401 below the llama.cpp Q6_K pipeline on 493 of 512. The cross-engine control cannot turn that into format parity or prove that bytes caused the earlier gap. The build still carries 2.306 GiB (13.5 %) less transformer body than Q6_K.

Two disciplines this exercise was run under, both worth more than the result: the payload was predicted at 23,035,310,148 B and measured at 23,035,310,148 B, zero error; and the registered 95 % interval [0.001175, 0.001601] missed — the measurement is 0.0000328 (2.0 %) above its upper bound, so all three estimators bracketed on the pessimistic side, with the role-share bound worst and the byte-law-at-Q6_K's-surplus form best, the opposite of what the surrogate's calibration implied. The published point prediction of ~0.0016 was within 2.1 % of measurement, the most accurate advance prediction this project has made — and the miss is printed beside it rather than dropped.

Related artifact: the measured answer to "what if one bit lower" is published as malaiwah/Qwen3.8-27B-EXL3-S16-V-research — a rejected sub-4-bit build at 0.045374 [0.041959, 0.049351], 4.39x this build's 0.010345, losing 512 of 512 contexts to it. It is published so the 16 GB no-go can be audited rather than taken on trust.

Correction, 2026-08-16 — the byte axis in the table above is not one axis, and every mixed comparison flattered us. A GGUF row is the whole file of a text-only artifact; our row is tensor payload of a multimodal tree that also carries an MTP draft. Read from each artifact's own tensor table, without downloading any payload (cross-candidate-byte-accounting.json):

All figures GiB. The embedding and head widths are not uniform across GGUF tiers, and the vision encoder is absent from every GGUF text file - it ships separately as mmproj-BF16.gguf, which no earlier comparison of ours counted. What that does to the four published claims, two against us and two for us:

1. 6 bits, Q6_K against hydrated: our sentence understated their byte spend roughly threefold. "+1.186 GiB more file" is +1.198 on tensors and +3.634 GiB of transformer body (23.1 % more than ours). Their fidelity win at 6 bits stands exactly as published - it is the price we mis-stated, in our own favour.
2. 6 bits, deployed: a multimodal Q6_K deployment is 22.18 GiB against our 20.13 GiB whole tree, so ours is 2.053 GiB smaller and needs no second file.
3. 5 bits, UD-Q5_K_XL against the context edition: this claim was wrong against us. We do not pay "0.445 GiB more" for the win - on transformer body they carry 2.148 GiB more (+14.4 %), and our deployed multimodal artifact is 0.422 GiB smaller.
4. 8 bits, Q8_0 against online K5/K6: the bodies agree to 1.7 % (24.526 against 24.119), the most format-comparable pair on the table, and they win it cleanly. That row is the reason the others are worth reading: this is not a table where every axis favours the author.

Rule from here on, and it is printed rather than footnoted: "at equal file bytes", "at equal tensor bytes", "at equal transformer body" and "as a deployed multimodal artifact" are four different claims, and whichever one a sentence means is written into the sentence. No fidelity number changes.

Q8_0 has the lowest measured complete-pipeline KL at 0.001087 for 27.05 GiB. Its capture engine differs from every vLLM row, so the 0.000507 BF16 control is a warning about confounding, not a resolution threshold or subtraction term. This build is nevertheless worse than every comparator by large observed margins, including same-engine official FP8.

Every GGUF point at or above 5 bits measures lower complete-pipeline KL than official FP8. That cross-engine observation makes the family's "below FP8" headline a weaker achievement than it sounds.

What this comparison does not settle. It is text-only teacher-forced fidelity on one shard of ten. It says nothing about serving 262,144 tokens with vision and MTP on a 32 GB card, which is where these artifacts actually differ and which is this build's own reason to exist, and llama.cpp KV-quant behaviour, prefill and decode speed are separate axes that were not measured here. The GGUF rows are a shard-0 ranking, not a paired per-context bootstrap against the ten-shard rows above, because those were welded from a different position count. Shard 0 is one tenth of the suite, and it is close to it: over all 10,480,640 positions the five vLLM means read 0.002760 / 0.003210 / 0.003509 / 0.005294 / 0.010604 — 1.9-2.9 % above these shard-0 values, ordering unchanged (receipts/kld5-10M-{hyd,k5k6,ctx,fp8,k4}.json). The GGUFs have no ten-shard equivalent; extending them is unrun.

One protocol objection, bounded rather than argued. llama-perplexity scores only the second half of each window, so every position it scores has at least 256 tokens of left context, while our suite scores from position 0. Re-scoring our own captures under that restriction lowers every candidate's mean by 1.3-2.1 % at a 256-token floor and 3.9-4.9 % second-half-only, uniformly enough to change no ordering — this build reads 0.010154 and 0.009876 respectively (receipts/scored-window-offset.json). The external protocol's scoring floor therefore explains at most about 5 % of any cross-protocol gap, and none of this build's distance from the rest of the table.

Cross-citation: the same three GGUFs under llama.cpp's own protocol

The rows above are those GGUFs on our axis. They have also been measured on theirs, run exactly as its authors run it, so the two can be cited side by side without either being converted into the other: llama-perplexity --kl-divergence on WikiText-2 raw test, n_ctx 512, 147,900 scored positions, KL(BF16 GGUF ‖ candidate) with both operands inside llama.cpp and each candidate's own output head inside the measured path, base Mean PPL 6.950230 ± 0.044933 (receipts/wikitext-kld-run-a.json; full protocol, delta by delta, in docs/35-external-protocol-comparability.md).

The ordering is identical, but no ratio or conversion is valid across the two protocols. Their run differs in corpus, scored-window geometry, candidate head and engine. Our run is additionally cross-engine for GGUF. The measured 0.000507 BF16 control proves confounding and cannot be subtracted. Correction, 2026-08-16: This paragraph used to call the output head "the large one". It is now measured on our own corpus and it is not: replaying each candidate through its own head over all 512 shard-0 contexts and 1,048,064 positions raises its mean by at most 5.28 % (hydrated 5.01 % of head-inclusive divergence, context 4.06 %, K4 1.17 %, unsloth NVFP4 2.64 %, and exactly 0 % for the official FP8 export, whose head is byte-identical to the shared one — the internal control), every interval excluding zero (receipts/head-attribution-v5.json). Scoring geometry is worth ≤4.9 % by the same kind of control. So the two protocol terms we have quantified are together far too small to explain a 1.1-1.6x level difference: the level gap is not decomposed, the leading unmeasured candidates being their 512-token English-encyclopedic windows against our 2,048-token five-strata ones, and the width of a GGUF's own output.weight, which is a different tensor from any head measured above.

Their harness's own floor, measured on our hardware instead of assumed. The Minimum KLD column is negative for all three — −0.000080, −0.000056, −0.000077, i.e. 5.6e-5 to 8.0e-5 — the uint16 16-nat log-probability encoding showing through rather than a candidate beating its own reference. The same term appears in the perplexity: 6.9525 in the capture log against 6.950230 in the scoring runs, identical weights on identical tokens, differing only by that stored round trip.

Tokenization is not part of the difference, and that is a measured null result. llama.cpp's GGUF BPE and our Hugging Face tokenizer produce bit-identical 297,194-token streams over this corpus — same int32 digest, no first divergence index — and the 296,960-token prefix that llama-perplexity actually scores is identical too (receipts/wikitext-kld-token-identity.json).

And one finding worth its own line: perplexity does not reproduce the KLD ordering. Q6_K has the smallest PPL delta of the three, +0.00079 against the 6.950230 base, while Q8_0 — the better quant by every divergence statistic, including a mean 2.5x lower and 0.9 points more top-1 agreement — is +0.00467. A quantization that shifts the distribution can shift it in the direction that happens to flatter a corpus mean, which is an argument for the metric this whole section is built on and against ranking quants by perplexity delta.

What this cross-citation cannot do is put this build on their axis: llama-perplexity cannot read an EXL3 checkpoint. The table at the top of this section, where every candidate is scored by one harness on one suite, stays the primary comparison.

Post-selection qualification

The v3 numbers in this card come from the suite that guided recipe selection. This is the v4 test that did not: 160 new contexts from 100 documents with zero intersection with the development suite (context token hashes 0/160, document names 0/100, content hashes 0/100), partitioned by whole source cluster, run once, with no recipe changed afterwards.

The original 42-context table used a fixed-stride character overlap scan. A later, offset-independent scan found exact 12-token calibration overlap in four qualification source documents. Applying the same conservative rule to every candidate — exclude every context from any source document with even one hit — leaves 36 contexts / 24 clusters:

The correction changes no ordering or paired win: the three EXL3 builds remain 47 / 41 / 32 % below FP8. The original 42-context figures and the candidate-independent correction are both preserved in docs/31. Absolute magnitudes remain suite-specific.

Public capability — MMLU-Pro, item-paired against BF16

70 MMLU-Pro questions, 14 official categories, 5 per category, official five-shot category prefixes, pinned TIGER-Lab/MMLU-Pro@b189ec765aa7ed75c8acfea42df31fdae71f97be, greedy, thinking at low reasoning effort, 5,120-token completion cap. The BF16 reference ran first; the plan and its acceptance rule were frozen before any result was seen. All six models answered the same 70 items in the same order through the same extractor, so every candidate row is paired item-by-item against the BF16 control.

This quant's verdict: 57/70 matches BF16 absolutely, and it misses the pre-registered bar at 88.1 %. Pass-outcome agreement is 66/70. Equal absolute scores with two regressions and two improvements is exactly why the measurement is item-paired: the totals match, the per-item behaviour does not, and the totals alone would have hidden both.

The pre-registered bar, and who clears it

The frozen plan accepts a candidate when BF16-pass retention has a Wilson 95 % lower bound at or above 0.90 and no category loses more than two BF16 passes. The category clause is met by all five candidates — the worst case is two passes in philosophy, for the hydrated build and online K5/K6 — so the retention lower bound is the only clause that ever fails.

Only the context edition clears the bar, at 90.7 %. This quant reads 88.1 % (55/57) and official Qwen/Qwen3.8-27B-FP8 reads exactly the same 88.1 % (55/57); the hydrated build and online K5/K6 read 85.6 %. Four of the five candidates, official FP8 included, are measured shortfalls, published as measured with nothing retuned after the fact.

No candidate here is shown to beat another. Every interval in the table overlaps every other interval, including the BF16 control's, so the matrix does not rank these models and this card does not claim it does. This quant is not shown to be worse than the context edition or better than the hydrated build or online K5/K6 on knowledge-and-reasoning tasks; on this suite it and official FP8 land on identical retention numbers. Note the contrast with the v5 tail result above, which is a distribution-fidelity measurement on 10,480,640 scored positions and does separate this build from FP8: that is a different question, on a different metric, at a different sample size, and it makes no capability claim.

Why a 70-item suite cannot certify this bar

With 57 BF16 passes as the paired denominator, 56/57 is the smallest count whose Wilson 95 % lower bound clears 0.90 (56/57 → 90.7 %; 55/57 → 88.1 %; 54/57 → 85.6 %). A single paired regression is therefore the entire budget, and no result that gives up two can pass, however sound the build. The suite has too few items to certify the bar it pre-registered, and at this size it separates nothing — the point applies to official FP8 exactly as it applies to this quant. Read the shortfall as a power limitation of a 70-item draw, not as evidence that any of these checkpoints is broken.

Four caveats, all load-bearing:

1. Four BF16 items hit the 5,120-token completion cap and are counted as failures under the plan's frozen addendum (finish_reason: length with no letter emitted), so the BF16 reference itself is depressed by the cap; this build hits the cap on four items too, and the per-model counts are in the table. The superseded 2,048-cap control, where BF16 lost 7/70 to truncation, is kept at receipts/public-capability-bf16-superseded-cap2048.json.
2. Exact-answer agreement is 0/70 for this build and for every other EXL3 candidate, and 1/70 for official FP8 (one math item, a 113-token answer both models pass). Long chains of thought differ token-wise on essentially every item, so the pairing is on pass/fail outcome, never on generated text.
3. This is a measured shortfall against the pre-registered bar. The plan required a Wilson 95 % lower bound at or above 90 % on BF16-pass retention; 88.1 % is below it. It is published as measured, with nothing retuned after the fact.
4. 70 items is small. The Wilson intervals are wide and mutually overlapping, so the full six-model matrix — now run for every candidate, not just this one — resolves no ordering. This is a first public, licence-compatible, item-paired benchmark, not a leaderboard claim. The honest next step is more items, which is the plan's own P1: HumanEval+/MBPP-style executable cases, IFEval-style constraint following, tool schemas, and a larger MMLU-Pro draw. No capability claim on this card graduates before that.

Receipts: plan, suite, and the six per-model runs linked in the table (receipts/public-capability-{bf16,ctx,k4,hyd,fp8,k5k6}.json); harness tools/public_capability.py, sweep runner tools/run_public_capability.sh. Every receipt carries the per-item raw request, raw response, extracted letter, gold letter and digests.

Downstream task retention — 40-task smoke suite (prior, narrower evidence)

This ran before the MMLU-Pro suite above and is kept unchanged. It is the narrower evidence: self-generated tasks with contract checks, not a public benchmark, and it says nothing about the knowledge-and-reasoning behaviour MMLU-Pro probes.

On 40 deterministic generated tasks (10 each arithmetic, executable builtins-only code, exact-list instruction following and tool-call schema), BF16 and every comparator scored 40/40. This build had zero regressions and matched BF16's exact final-answer text on 33/40; all seven differing answers still passed their contracts. Wilson 95 % lower bound is 91.2 %. This is a transparent smoke suite, not a public leaderboard; full responses are in the run receipt; its extracted-value agreement field is superseded by the strict rescore.

Serving

Requires the Gilded Gnosis vLLM fork — the EXL3 checkpoint loader, the B12X Trellis kernels and the exl3-b6 online overlay are not in upstream vLLM. The public image is:

voipmonitor/vllm:gilded-gnosis-v20-vllm4d006a4-b12xcd3ce19-fi1ac6942-cu132-20260810-r34
registry digest sha256:820181fbbc975cd5291c411cda9771d58fecee1636d916f508f47230df20592b

Its launcher scripts only dispatch GLM-5.2 and DeepSeek families, so call vllm serve directly:

docker run --rm --gpus '"device=0"' --ipc host -p 127.0.0.1:8000:8000 \
  -v /models:/models:ro -v /cache:/cache \
  -e VLLM_EXL3_ONLINE_TRELLIS_BITS=6 \
  -e VLLM_EXL3_ONLINE_CACHE_DIR=/cache/exl3-online \
  -e VLLM_EXL3_ONLINE_CACHE_MODE=readwrite \
  --entrypoint /opt/venv/bin/vllm \
  voipmonitor/vllm@sha256:820181fbbc975cd5291c411cda9771d58fecee1636d916f508f47230df20592b \
  serve /models/Qwen3.8-27B-K4 \
    --served-model-name qwen38-k4 \
    --quantization exl3 \
    --enforce-eager \
    --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*visual\\..*","re:.*in_proj_a$","re:.*in_proj_b$","re:.*in_proj_ba$","re:.*mtp\\..*","lm_head"]}' \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.85 \
    --max-num-seqs 4 \
    --host 0.0.0.0 --port 8000

The container listens on all interfaces internally, but Docker publishes the port to host loopback only. For remote clients, keep that binding and put an authenticated TLS proxy in front; do not expose this unauthenticated generation endpoint directly.

Four flags are load-bearing:

1. --quantization exl3 is mandatory. Auto-detection only fires for the GLM-5.2 r7_routed_experts / hybrid_tr3_tail metadata; a dense tensor_storage checkpoint is not auto-detected.
2. --enforce-eager is mandatory. The loader refuses non-eager execution for any checkpoint without rank-sliced metadata, because exl3_gemm autotunes with timing launches. Expect no CUDA graphs.
3. The ignore list is mandatory and its anchoring is subtle. The overlay claims every BF16 LinearBase not present in tensor_storage — including the vision tower and the MTP head. The prefixes it matches have no leading model., so re:.*\.visual\..* (dot before visual) silently fails to match while re:.*visual\..* works. With the wrong pattern the vision tower is claimed and startup crashes (ValueError: MXFP8 requires input_size_per_partition (4304) to be divisible by 32), reported upstream as local-inference-lab/vllm#311 with a verified fix in PR #312, which degrades those shards to BF16 with a warning instead of aborting.
4. VLLM_EXL3_ONLINE_TRELLIS_BITS=6 is what turns the overlay from MXFP8 into K6. Point VLLM_EXL3_ONLINE_CACHE_DIR at persistent storage: the first load encodes 208 attention projections (~16 min on one GPU here) and later loads reuse the cache.

Prefix caching is on in the recipe below, and it is on here and not everywhere. At an 8,192-token window the KV pool is roughly thirty times the window, so a single request comes nowhere near the pool ceiling and the failures that stopped the native-context profile cannot occur. That is not an argument, it is the reason this recipe was measured separately: it starts healthy on the promoted image with --enable-prefix-caching --mamba-cache-mode align, answers a text and an image request exactly, and reports enable_prefix_caching: True in the engine banner (receipts/production-image.json). The context edition's native 262,144-token recipe does not enable it, and its card explains why (receipts/qualification-5090-apc.json).

The release unit moved on 2026-08-16, and #51113 is why. Upstream vLLM #51113 (mamba align prefill-chunk splitting: a chunk that ends mid-block leaves its slot holding a short state, which a later chunk then publishes anyway — wrong tokens, HTTP 200, no crash) merged 2026-08-06, after the pinned public image was built, and is still absent from it and from fork head fa033bd4e. Cherry-picks were requested upstream on 2026-08-16 (issue #392, PR #393). Until they land it is carried as tools/vllm-mamba-align-scheduler.py (sha256 b431c106…), and the image this project serves is now the four-module localhost/vllm:gg-r34-patched-apc, manifest sha256:16a936b877b90f…, promoted from the three-module localhost/vllm:gg-r34-patched (sha256:6eca4c693f01b6…) (receipts/production-image.json). With prefix caching off the two images are not merely similar but behaviourally identical, because the added module's changed function is unreachable unless mamba_cache_mode is align — so the earlier hardware qualification carries over unchanged. One warning if you inspect the image yourself: its build-time label io.malaiwah.image.qualified still reads false and is superseded by the receipts named here — it was written before the image could possibly have been qualified, and correcting it would add a layer and change the very digest that was measured. That digest is local to the build host, so the recipes here reproduce its content with sha256-verified read-only mounts over the pullable public base.

What prefix caching buys. On disjoint documents, so the cold case is genuinely cold: a 32,842-token prefix went 12.07 s cold → 1.04 s warm (11.6×, 2,442 of 32,842 prompt tokens recomputed, 92.6 % hit rate) and a 131,146-token prefix went 67.60 s → 2.31 s (29.3×, 3,146 of 131,146 recomputed, 97.6 %); a 38-request schedule ran 84.0 s with the cache against 144.4 s without (receipts/apc-poison-repro.json). And what we actually know about its safety. Correctness was probed adversarially before any of this shipped: seven freshly started servers, 38 requests each, 266 scored requests, nested token prefixes so later requests hit blocks published by earlier ones, and no prompt length a multiple of the measured 1,600-token mamba block, so prefill chunks end mid-block by construction — zero corrupted responses, zero wrong answers, zero acceptance collapses, on the unpatched image as well as the patched one. Thresholds were committed before the first server started; the worst repeated block was 15 characters against an 80-character threshold, with no U+FFFD anywhere. Greedy chosen-logprob drift with the cache on is 0.1063 mean absolute against a measured run-to-run floor of 0.0823 — drift, never an answer change. So the module is carried as insurance backed by upstream's own regression file — 14 failed / 6 passed against the vendored scheduler, 20 passed against this one — and not by a reproduction of our own: we tried hard to reproduce the reported corruption and could not (receipts/mamba-align-defect.json).

LMCache is unmeasured by us. It is not part of any recipe on this card, this project has never run it, and it is the outstanding suspect in the one user report of prefix-cache corruption we have. Nothing here says LMCache is safe; the evidence above covers vLLM's own prefix cache and nothing else.

#51812 is now recommended if you serve this recipe concurrently with MTP, and it stays an overlay. Upstream #51812 (Qwen GDN speculative gate ordering: the vendored code gathers the speculative Q/K/V rows but hands the recurrent update the ungathered a/b gate tensors, so gate row i can belong to a different token than Q/K/V row i) merged 2026-08-11 and is absent from the promoted image. Whether that path is ever entered is no longer an argument — it was counted. A CPU-only instrument over the engine's GDN metadata builder ran the three flags this card publishes for prefix caching (--max-model-len 8192, --enable-prefix-caching --mamba-cache-mode align, --gpu-memory-utilization 0.85) at eight concurrent streams, with MTP-3, fp8 KV and --max-num-batched-tokens 2048 supplied by the arm, and the defective path was entered: three events in 5,825 metadata builds, 0.515 per thousand, over 468 requests with zero errors and a prefix-cache hit rate reaching 50.1 % (receipts/gdn-gate-concurrency.json). The mechanism, in one sentence you can act on: at eight streams with prefix caching on, a short non-speculative request can land between speculative ones in the same batch, and the unpatched gather then misaligns the gates. All three firings had one shape — six speculative decodes plus one non-speculative request, that request beginning at token index 20 and displacing one four-token speculative decode, whose four gate rows were read from the wrong tokens. Two boundaries on that number. The same instrument saw zero events in 3,329 builds at eight streams with prefix caching off (below 0.90 per thousand, 95 % upper bound) and zero in 8,065 builds at a quarter of the token budget, so it is the cache path that opens this rather than concurrency alone. And the arm served the context edition's weights: the counter reads only the scheduler's host-side arrays, so what it measures is a function of flags and traffic, with the checkpoint entering only through the size of the KV pool.

Two conditions gate all of this, and the command below is missing one of them. The changed code runs only when a batch carries speculative tokens at all — the vendored file takes its speculative branch only when spec_sequence_masks is not None — so with no --speculative-config, as the command below ships, the module is a no-op by construction; and at --max-num-seqs 1 no mixed batch can form either. Add MTP to this recipe and the measurement above is what applies. The command below ships --max-num-seqs 4, four streams rather than the eight that were measured, where the same mechanism exists and its rate is unmeasured. When both conditions hold, mount it read-only: tools/vllm-qwen-gdn-spec-gates.py (sha256 7cd3f5fe763b621048af4817951a841d99c8b700d9a56ded27ccaca5a56ccbe0) over /opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py. It is diff-identical to upstream's eight changed lines and py_compile-clean under the image's Python 3.12.3, and it is an overlay deliberately not part of the qualified digest: sha256:16a936b877b90f… is what was qualified, and a reachability count is not evidence that would survive a re-qualification, so it is mounted over the vendored file rather than promoted into the image.

The effect on answers was not measured, and measuring it was declined on resolution grounds. Three events in 5,825 builds cannot move a statistic whose run-to-run floor is 0.0823 mean absolute chosen-logprob error against a per-event effect of 0.002755 — a noise floor about thirty times the size of one event — so an A/B would have returned its own noise and was deliberately not run. Nothing here claims the overlay changes an answer, or that it does not. The recommendation follows a rule fixed before the GPU window opened instead: a nonzero rate in a regime we ship gets the free fix, because a silently miscomputed forward pass gives the operator no signal at all. Traffic in that arm was adversarial by design — eight speculative streams plus injected short prompts and full-cache-hit repeats — so 0.515 per thousand is an upper bound on a shipped regime, not a forecast for your workload (receipts/gdn-spec-gate-defect.json is the source-level defect analysis).

Scope on this build, stated narrowly. What was measured on this profile is that the recipe starts healthy on the promoted image with the cache on, answers a text and an image request exactly, and reports the cache enabled in its banner. That is a serving smoke, not a gate suite: this window has no long-needle, combined-image or decode-dispersion gate of its own, and the 11.6× and 29.3× reuse figures above were measured on the context edition's much longer prompts, not on 8,192-token ones. Expect the shape of the win — recomputing only what changed — rather than those multiples.

Enabling it on this profile is one extra read-only mount and two flags:

set -euo pipefail
git clone https://github.com/malaiwah/qwen38-27b-exl3 && cd qwen38-27b-exl3
cat <<'SHA256' | sha256sum -c -
b431c1066dfee3ed56bfa7e71cc8606f9afadc300f22d7fc542c43835d1b22bf  tools/vllm-mamba-align-scheduler.py
SHA256
SCHED=$PWD/tools/vllm-mamba-align-scheduler.py

docker run --rm --gpus '"device=0"' --ipc host -p 127.0.0.1:8000:8000 \
  -v /models:/models:ro -v /cache:/cache \
  -v "$SCHED:/opt/venv/lib/python3.12/site-packages/vllm/v1/core/sched/scheduler.py:ro" \
  -e VLLM_EXL3_ONLINE_TRELLIS_BITS=6 \
  -e VLLM_EXL3_ONLINE_CACHE_DIR=/cache/exl3-online \
  -e VLLM_EXL3_ONLINE_CACHE_MODE=readwrite \
  --entrypoint /opt/venv/bin/vllm \
  voipmonitor/vllm@sha256:820181fbbc975cd5291c411cda9771d58fecee1636d916f508f47230df20592b \
  serve /models/Qwen3.8-27B-K4 \
    --served-model-name qwen38-k4 \
    --quantization exl3 \
    --enforce-eager \
    --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*visual\\..*","re:.*in_proj_a$","re:.*in_proj_b$","re:.*in_proj_ba$","re:.*mtp\\..*","lm_head"]}' \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.85 \
    --max-num-seqs 4 \
    --enable-prefix-caching --mamba-cache-mode align \
    --host 0.0.0.0 --port 8000

Generation defaults from upstream generation_config.json: temperature 1.0, top_p 0.95, top_k 20. Thinking control is upstream's chat_template_kwargs: {"enable_thinking": false} or {"reasoning_effort": "xhigh"|"medium"|"low"} (upstream raises on high). The chat template, tokenizer, preprocessor configs and vocabulary (248320, untied head) are upstream's, unmodified.

Context: 262144 native, verified here only to 8,192. Upstream's 1M procedure is static YaRN (nested rope_parameters with rope_type: yarn, factor: 4.0, original_max_position_embeddings: 262144, VLLM_ALLOW_LONG_MAX_MODEL_LEN=1, --max-model-len 1000000), not a bare max_position_embeddings bump, and Qwen warns it costs short-context quality. Untested on this runtime.

KV-cache dtype: fp8 is the family's measured default

The recipe above leaves the KV cache at the engine default, as the Tradeoffs section notes; the family's qualified long-context profiles pin fp8, and that default is now measured rather than assumed. A five-arm sweep on the physical RTX 5090 served the context edition — same engine, same flag surface — at its native-window profile with the KV dtype the only deliberate flag change: no arm beat fp8 on native-or-beyond context on 32 GB with retrieval intact (receipts/kv-dtype-sweep-5090.json, decision record docs/38-kv-dtype-sweep.md). The engine derives the attention backend from the KV dtype, so each arm is measured as it actually serves:

The per-token-head family is measured, not assumed: int8 and fp8 per-token-head each dominate fp8 on both capacity and closeness to the bfloat16-KV reference, but each pays 3.0× prefill because TRITON_ATTN is the only backend on this fork that accepts per-token-head scales — and the capacity edge is TRITON_ATTN's smaller CUDA-graph pool (0.06 against 0.45 GiB), not cheaper bytes: those arms cost more per token than fp8 (35,360 against 34,816 B/token). int4_per_token_head is the real capacity lever — 502,667 tokens, 1.92× concurrency — at two named prices: 3.6× fp8's distributional error and 2.78× its prefill. nvfp4 and nvfp4_ds_mla do not start: no attention backend on this fork advertises nvfp4 for a non-MLA decoder — all five candidates answer kv_cache_dtype not supported — and the GLM-5.2-serves-nvfp4 precedent is the owner's claim about a different model, unverified here. The fidelity column is a bfloat16-KV-reference probe at a 98,304-token context — truncated top-20 KL over 70–173 paired greedy positions, a lower bound; it is not the v5 KLD and must never be differenced against any published KLD figure. Retrieval was 44/44 exact across the five arms, 4-bit included — retrieval is not fidelity, which is exactly why the KL column exists.

Reconstruct-scratch arena: +17,874 KV tokens, and it stays an overlay

A 2-hunk fork patch to exl3.py buys 17,874 more KV tokens, measured, and it is an opt-in overlay rather than part of any qualified digest. The pinned r34 image keeps one persistent fp16 prefill-reconstruct scratch per weight geometry — 790 MiB across the eight geometries that allocate at the qualified long-context profile (the head's 5120×32768 chunk needs ≥128 sampled logit rows and never triggered). The patch (overlay tools/vllm-exl3-scratch-arena.py, sha256 9aba06ebf60ca7665c0513752387c349240ab85e1ebc44d6ce8137ef157b6c15; fork PR local-inference-lab/vllm#397) shares one grow-to-max arena per device instead, sized by the largest live geometry (170 MiB), because each reconstruct is written and consumed inside one eager call on one stream. The kernels see identical operands — same shapes, strides and dtypes.

Measured on the physical RTX 5090 A/B: engine-reported KV pool 265,122 → 282,996 tokens (+17,874, +6.7 %, 9.28 → 9.88 GiB ≈ +0.60 GiB), reproduced identically across two server starts per arm, the arena's own growth log ending at exactly the predicted 170 MiB (receipts/scratch-arena.json). That A/B served the context edition at its qualified 262,144-token profile — same engine, same flag surface — and the scratch geometries are shape-derived, so the mechanism and the 170 MiB arena size are identical on this build while the pool figures above are the context edition's. Fidelity was gated rather than assumed: the 30-case deterministic vision suite returned byte-identical answers on both arms (24/30 each, equal to the rank-1 qualification reference), a full-window needle (258,925 tokens, depth 0.5) retrieved exactly, and decode did not regress (109.2–109.7 against 108.5–108.9 tok/s over three warmed C1 runs). Read the byte-identity claim narrowly — it covers that deterministic probe set, because the control shows two restarts of the unpatched baseline differ on 7 of 8 long greedy continuations (exl3_gemm autotunes kernel configs by measured time per process), so this stack is not restart-deterministic on long greedy text with or without the patch, and every cross-restart pairing is 7 DIFF / 1 MATCH either way.

Like #51812, it is an overlay deliberately not part of the qualified digest: the pinned digest is what was qualified, a larger KV pool is not evidence that would survive a re-qualification, so it is mounted read-only over the vendored file rather than promoted into the image — -v tools/vllm-exl3-scratch-arena.py:/opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/exl3.py:ro. The static prediction had been +620 MiB / +18.7k tokens; the measured gain is 95.7 % of it, and the measured number is the one to quote. On the 24 GB class the same bytes put the published 24,576-token window at 42,450 raw token headroom, supporting 40,960 at the next window step — arithmetic only, pending a 24 GB-class boot (docs/34-vram-class-profiles.md §10.2).

Chat template

This repo ships chat_template.jinja byte-identical to Qwen/Qwen3.8-27B (sha256 c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041), and the same bytes again in the chat_template key of tokenizer_config.json. Do not replace one without the other: under transformers 5.15.0 the .jinja file takes priority, so editing only tokenizer_config.json is a no-op. To override, pass --chat-template <file>.

Three upstream-template restrictions to code against. All three are Qwen's, unchanged by us, and all three surface as HTTP 400:

• reasoning_effort accepts only xhigh (default), medium and low - measured against a live endpoint, correcting an earlier revision of this card that also listed none. The server answers HTTP 400 Unexpected reasoning effort none. Supported types are xhigh (default), medium, and low. for none, exactly as it does for high, minimal and max - all of which vLLM's OpenAI surface advertises and this template rejects. If your client hard-codes high, serve with --default-chat-template-kwargs.reasoning_effort=xhigh.

• What the ladder costs, measured rather than described. Same prompt, temperature 0, 32,768-token budget so no arm is truncated, two repeats each, on the qualified 1x endpoint (receipts/reasoning-effort-1x.json):

  reasoning_effort	completion tokens	reasoning chars	latency	finish	correct algorithm
  (unset)	18,085	56,955	168.9 s	stop	2/2
  xhigh (default)	18,085	56,955	168.8 s	stop	2/2
  medium	3,436	4,767	26.2 s	stop	2/2
  low	3,517	4,606	27.0 s	stop	2/2

  Three things follow. Unset is byte-identical to xhigh, which is how the default is now measured rather than assumed. xhigh costs 5.3x the output tokens and 6.4x the wall clock of medium for the same correct answer on this task - the single largest cost knob on the client side of this card, and the one to reach for first in any agent loop bounded by a per-task timeout rather than by quality. And low is not cheaper than medium (3,517 against 3,436 tokens): the bottom of the ladder is flat, so medium is the useful floor. Read narrowly: one prompt, one shape, an exact-match check on whether the right algorithm is named - it establishes token cost, not a quality ordering.

• Exactly one system message, first. Two system messages 400 with System message must be at the beginning. (vllm#41114, open; fix PR vllm#44505 closed unmerged). Merge them client-side. A developer role is safe - vLLM folds and consolidates it (vllm#43590).

• Content blocks must be text, image, image_url or video. Anything else, including an Anthropic tool_result carrying a tool_reference item on /v1/messages, 400s with Unexpected item type in content. (vllm#52489, open).

• Only the leading system message is accepted. A second or mid-conversation system role raises. Clients must merge system instructions into the first system message before rendering history.

• Tool-call argument values are an unescaped XML boundary. Before replaying assistant history, reject values containing literal </parameter> or <parameter= delimiters. Treat replayed calls and tool output as untrusted data; the template cannot escape those sequences.

Echoing reasoning_content back on assistant history turns is what buys a full prefix-cache hit: measured 100 % prefix reuse when the client returns it, 94.7 % at ten turns when it does not. Community "fixed" Qwen templates are not recommended here: measured against --tool-call-parser qwen3_coder, qwen3.8-froggeric-v22 renders a tool call whose arguments the parser recovers as {}. Details in docs/39-chat-template-audit.md and receipts/chat-template-audit.json.

Verification

All measured on 1x RTX PRO 6000 Blackwell Server Edition (SM120, 96 GB), driver 595.58.03, TP1, with the r34 image above.

Loads and serves. Engine reports quantization=exl3, online K6 encoding for every attention projection (proxy error ~3.2e-4 per projection), then:

Actual usage is 17.89 GiB for weight, 2.33 GiB for peak activation,
0.26 GiB for non-torch memory, and 0.0 GiB for CUDAGraph memory
GPU KV cache size: 736,109 tokens   (--max-model-len 8192, --gpu-memory-utilization 0.85)

Text is coherent. "Name the three primary additive colors, comma separated." with enable_thinking: false → Red, Green, Blue (1.6 s, greedy).

Vision works. A 96x96 PNG, left half pure red, right half pure blue, with "Name the left colour then the right colour, comma separated." → red, blue.

Quantization error, per tensor, from the conversion log (LDLQ proxy error): down_proj is consistently the worst projection in every layer — about 2.5e-3 versus 1.1e-3 for gate_proj and 1.0e-3 for in_proj_qkv. Whole-block figures: rfn ~0.0155, sqnr ~36.4 dB.

Distribution fidelity — v3 protocol, held-out corpus (prior receipt)

Superseded as the headline by the v5 suite above, which scores 10,480,640 positions against this section's 278,392. Kept unchanged, including its own correction history.

The figure's four points and four top-1 values are the overlap-corrected 127-context subset; the table below is the original full 136-context receipt, which is why 94.48 / 96.18 / 90.49 % there reads 94.50 / 96.22 / 90.53 % here.

136 analysis contexts x 2047 positions = 278,392 scored positions from separately sourced Gutenberg, arXiv, Wikipedia and CPython documents. The original fixed-stride 160-character scan reported zero calibration hits; a later all-position 12-token scan found exact overlap in 2/41 source documents. Exact full-vocabulary two-pass KL(BF16 reference || candidate) through one shared BF16 LM head, float32 within each vocabulary chunk and float64 across chunks, source-cluster bootstrap.

The weight column prints the measured resident weights of docs/22-results-iteration-2.md — 28.51 GiB = 30.61 GB for FP8, 17.89 GiB = 19.21 GB here, 21.34 GiB = 22.91 GB for unsloth NVFP4. Earlier revisions of this card printed 30.9 / 19.2 / 23.4 GB in this column, the older docs/18 presentation of the same three measurements. The KLD, interval, median, p99.9, JSD and top-1 columns are unchanged.

Overlap-corrected subset: conservatively removing all nine analysis contexts from either affected source document gives K4 0.029679, FP8 0.012798, and NVFP4 0.092727 over 127 contexts. The ranking and every conclusion survive; the table above is retained as the original full-suite receipt, not described as contamination-free.

Never quote the v3 NVFP4 number as our current one. NVFP4 reads 0.092727 on this corrected v3 subset and 0.030115 on v5 shard 0 — same checkpoint, same revision, same flags, same shared-head protocol. That gap is suite hardness, and it was measured for all six candidates rather than argued: v3-corrected ÷ v5 shard 0 is 2.4625x official FP8, 2.5293x online K5/K6, 2.6564x hydrated, 2.7505x the context edition, 2.8688x this build and 3.0791x NVFP4 — a band spanning 1.2504x end to end, with the ordering identical in both suites (receipts/nvfp4-v5-measurement.json, block suite_comparability_v3_vs_v5). A band of factors and not one factor is exactly why no conversion between the suites exists: the ordering carries across, an absolute value never does, and a v3 number must never appear in the same sentence as a v5 number.

Paired over the same contexts: -0.064242 versus NVFP4 (95 % CI [-0.08621, -0.04611], 136/136 contexts ours) and +0.017611 versus FP8 (95 % CI [0.01256, 0.02368], 136/136 contexts FP8).

An earlier version of this card reported better numbers on a contaminated suite. The previous corpus was exllamav3's own calibration data — the text this quant was tuned on, while the NVFP4 and FP8 candidates were calibrated elsewhere. Re-measuring on held-out text moved ours from 0.026231 to 0.030736 (+17 %), NVFP4's from 0.073006 to 0.094978, and FP8's from 0.019309 to 0.013126 (-32 %). These are the honest numbers; the correction is documented in the companion repo.

Controls shipped with the dataset: runtime-repeat noise floor 0.000000 across three captures of the same runtime (this runtime is bit-deterministic, so every difference above is far outside noise); harness self-check 0.000000; CUDA-graph and a harness self-check of 0.000000. A third control, "CUDA-graph parity 0.000000", is withdrawn: it captured a prefill forward, and FULL_DECODE_ONLY captures no prefill graph, so it could not have measured the decode path. Re-measured on real decode steps, graph and eager agree on 24/32 greedy 32-token sequences (mean |Δ logprob| 0.0118) and unquantised BF16 on the same build drifts identically (24/32, 0.0128), so the drift is a property of CUDA graphs here rather than of the quantisation. Replay qualification is the weak link at 5.83e-04 mean, re-derived inside the v5 suite (32 contexts, 65,504 positions, [5.15e-04, 6.64e-04], superseding the six-context v3 figure of 6.54e-04), KL(live || replayed) — 2 % of this candidate's KLD and 4 % of the gap to FP8, so no ranking depends on it, but differences below ~1e-3 are not resolvable with these artifacts.

Head attribution: the K6 lm_head is nearly free

Replaying the identical stored hidden states through the BF16 head and through the reconstructed K6 head (exllamav3's own reconstruct_had_slice, so it is the exact serving matrix) isolates head error from body error:

The K6 head adds 6.78e-05 on top of the body (95 % CI [4.63e-05, 9.01e-05]), i.e. 0.26 % of total divergence. Contrary to the common assumption that lm_head is highly quantization-sensitive, at 6 bits on this model it is not worth spending 1.6 GB to promote it to BF16 — that budget belongs to the MLP stack, which owns the rest of the error.

Single-window KLD, v1 protocol (kept for continuity)

This was the first measurement; the v2 protocol above supersedes it.

Teacher-forced KLD, full vocabulary

One frozen 2048-token window (exllamav3's bundled wiki.utf8, first 2048 tokens), 2047 scored positions, KL(BF16 teacher || candidate) across the entire 248320-token vocabulary with no top-k, 3 repeats, --kv-cache-dtype auto pinned for every candidate, same teacher logits file for all of them. Protocol and statistics follow the published Gilded Gnosis harness (rtx6kpro:scripts/glm52_exl3_shared_h_kld.py): the headline value is the mean of the per-run means and run SD is the sample SD across those means.

This quant is 2.7x closer to the BF16 teacher than the same-generation NVFP4 checkpoint, while holding 3.70 GB less resident weight. That is the whole point of the recipe: Trellis K4 spends 4.004 bpw where NVFP4 spends 4.50, and the savings buy K6 attention instead of FP8.

run SD = 0 for both candidates means the three repeats were bit-identical — expected for the eager, max_num_seqs=1, prefix-caching-disabled configuration, and a useful signal that the online-K6 cache reloads deterministically.

For scale, this project uses project-local, unvalidated descriptors (<0.01 near-lossless, 0.01-0.05 good, 0.05-0.1 noticeable, >0.1 significant); they are not an external standard and do not transfer across models, corpora or tokenizers. This quant sits in the "good" band; the NVFP4 control sits in "noticeable".

Still measuring on the same window and teacher: this checkpoint, overlay off (attention stays BF16 in VRAM).

Throughput — with CUDA graphs

Same GPU, --max-num-seqs 8, greedy, ignore_eos, 256 output tokens, warmup discarded.

Graphs are worth +92 % / +84 % / +98 % here — roughly nine times what they buy the BF16 model (+8-11 %), because eager EXL3 pays per-call dispatch on 193 quantized matmuls. With graphs this quant is both the smallest and the fastest option measured; the earlier claim that distribution parity against eager is exact has been withdrawn (it measured prefill, which FULL_DECODE_ONLY never captures) and replaced by a real decode probe: 24/32 exact sequences, with a BF16 control showing the same 24/32.

Graphs need the patch in local-inference-lab/vllm#312's sibling (autotune priming, filed separately) plus:

-e VLLM_EXL3_GRAPH_DECODE=1 ... --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}'

Without that patch the loader refuses non-eager execution and you must pass --enforce-eager, which costs the throughput above.

Concurrent serving: speculative depth is a concurrency-dependent choice

Measured on the user's own physical GeForce RTX 5090 (32,607 MiB, driver 610.57.04) on the immutable production image localhost/vllm:gg-r34-patched (sha256:6eca4c69…) with no source bind mounts, on the context edition rather than this checkpoint — 11 configurations, identical frozen token-id prompts, three warmed repeats (receipts/perf-sweep-5090.json, decision record docs/36-performance-levers-5090.md). No figure in this subsection may be differenced against the rental RTX PRO 6000 table above, per the receipt's own rule; it is carried here because the finding is a serving recommendation that applies to every build in the family.

The decision metric is accepted tokens per step ÷ step time, never acceptance rate. The reference row, MTP depth 3 at --max-num-seqs 8, reads 82.94 / 263.12 / 313.28 tok/s aggregate at 1 / 4 / 8 concurrent streams with step times 25.72 / 30.64 / 49.52 ms. At one stream, MTP depth 3 wins: 2.1429 accepted tokens per step over a 25.72 ms step = 83.31 per-request tok/s, against depth 1's 1.6558 over 22.29 ms = 74.28. At eight streams the trade reverses decisively: num_speculative_tokens=1 drops accepted tokens per step by 22.98 % (2.1753 → 1.6754) while step time falls 37.25 % (49.52 → 31.07 ms), so aggregate throughput rises 30.67 %, 313.28 → 409.35 tok/s (+22.75 % per request), and it holds 10,911 more KV tokens (283,481 against 272,570) — the only row whose needle ran at the full 261,794 tokens, retrieved exactly, with the 30-case image suite unchanged at 24/30. Depth 2 is dominated at both ends. So: keep depth 3 for interactive single-stream use, set depth 1 when the deployment really runs concurrent streams, and note that at concurrency 4 the choice does not matter (±2.6 %). One constraint travels with the 8-stream matrix: at 262,144 tokens it only starts at --gpu-memory-utilization 0.97, which is a text-only profile because a large image OOMs in the vision tower there, so a vision-capable deployment stays at one sequence. The crossover exists because each drafted token costs a fixed slice of step time whose acceptance does not improve with batch size, while a wider batch already fills the step: past some concurrency the cheaper step buys more than the deeper draft.

What that means for the recipe above, which does not enable speculative decoding: if you add --speculative-config '{"method":"mtp","num_speculative_tokens":N}', choose N = 3 for interactive single-stream serving and N = 1 once the deployment really runs eight concurrent streams. That is the only lever in the sweep worth changing.

Closed avenues, so nobody re-runs them. --attention-backend FLASHINFER is a no-op: the engine already auto-selects FlashInfer on SM120 with fp8 KV and head_size 256. Forcing TRITON_ATTN, which the live K4 service does, is up to 5.5 % worse on step time at eight streams and its apparent +5.5 % gain at temperature 0 becomes −7.3 % at temperature 0.6, so it is acceptance noise rather than throughput. custom_ops:["all"] is 2.2-5.2 % worse on step time and is not bit-exact. Both dynamic speculative-decoding knobs are structurally unusable here: either one downgrades cudagraph_mode from FULL_DECODE_ONLY to PIECEWISE, which Exl3Config refuses, so the server does not start — and forced eager, the only form that runs, loses 48 % of decode. Dynamic speculative depth is no longer closed — it was a fixable bug, now fixed and measured (2026-08-16). The cudaErrorIllegalAddress that made VLLM_USE_V2_MODEL_RUNNER=1 unusable was a FlashInfer gate that admitted persistent (CUDA-graph) decode wrappers only for q_len == 1 + num_speculative_tokens, while the speculator's draft steps run q_len == 1; the draft-decode graph was therefore captured and replayed on the dynamic wrapper, whose plan buffers move on every call. Keying wrappers by the shape capture actually planned fixes it (local-inference-lab/vllm#398, closes #396; root cause proven with an instrumented capture-versus-replay address log plus a control that changes nothing except pinning the capture-time buffers alive). With the fix, one server running the per-batch-size schedule (depth 3 at batch 1-2, depth 1 at batch 3-8) measures 87.4 tok/s at C1 and 416.3 tok/s at C8, +38.0 % aggregate decode at C8 against an MRV1 baseline measured in the same window (301.6 tok/s) with C1 held — depth-1 throughput at eight streams without giving up depth-3 latency at one. Greedy outputs repeat token-for-token within an engine process in all four arms, and acceptance does not drop, so the speed is not bought with looser verification. The constraint that decides whether you want it: at --gpu-memory-utilization 0.97 the V2 runner leaves 58.56 MiB free and the EXL3 prefill reconstruct OOMs on the first 2,048-token prefill (both arms, schedule off and on), so the win was measured at 131,072 context and 0.95 utilisation — 234,256 KV tokens against 272,570 on this card's published profile, −14.1 %. Serving the published 262,144/0.97 profile under the V2 runner on a 32 GB card is not demonstrated. Cross-restart bit-exactness remains unclaimable on this stack (exl3_gemm autotune is per process). The static depths, 3 single-stream and 1 at eight streams, remain the published recipe until that KV concession is either accepted or removed (receipts/v2-runner-depth-schedule.json). Prefill did not move on any lever (3,374.4 tok/s at 2,048 and 3,255.4 at 6,144 prompt tokens for the reference row, no graph-decode arm more than 4.0 % away), so the prefill deficit is structural rather than untuned.

One reconciliation, because both numbers are published: the sweep's 82.94 tok/s at one stream sits below the context edition's qualification median of 107.56 tok/s purely because of acceptance, not speed — 2.14 accepted tokens per step here against 2.69 there, since these frozen prompts are literary prose — while step time agrees to 2.6 % (25.72 ms against 25.05 implied). The two measurements are consistent.

Tradeoffs, stated plainly

• The download is 28.31 GB for a 19.21 GB resident model. Attention ships BF16 so the runtime can encode it at K6 (and, later, at another width) instead of being locked to a serialized choice. If you want the download to equal the resident weight figure, the v-serialized-k6 variant is the one to ask for.
• CUDA graphs need a patched loader (see the throughput section). Unpatched, the loader refuses non-eager execution and you lose 46-50 % of decode throughput (local-inference-lab/vllm#311 tracks the surrounding overlay work; the graph guard itself is next on the list). Decode is 58-60 % of the NVFP4 checkpoint's, dominated by the GEMM kernel rather than by graphs.
• First load pays the K6 encode (~16 min here) unless the cache directory is warm.
• One runtime. This checkpoint does not load in upstream vLLM, SGLang, transformers, TensorRT-LLM or llama.cpp. exllamav3 itself can read the serialized K4/K6 halves, but it will not perform the runtime K6 encode of the BF16 attention.
• KV cache is left at engine default (auto); both NVFP4 references quietly ship FP8 KV schemes. Pin --kv-cache-dtype explicitly if you are comparing.

Reproducing this quant

# 1. Convert everything at K4 (vision left BF16, head K6, mcg codebook).
python convert.py -i Qwen3.8-27B -o qwen38-k4 -w wd-k4 \
  -b 4 -hb 6 -mb 4 -vb 16 -cb mcg -d 0        # exllamav3 1.4.2 @ 5f3c537

# 2. Splice BF16 attention + MTP back over the K4 output; the converter cannot
#    emit BF16 for a decoder linear (load_fp16 forces float2half).
python splice_bf16_attn.py -q qwen38-k4 -s Qwen3.8-27B -o Qwen3.8-27B-K4

# 3. Regenerate metadata so tensor_storage describes the mix.
python util/add_safetensors_index.py -m Qwen3.8-27B-K4 --force
python util/add_quant_config.py -m Qwen3.8-27B-K4

splice_bf16_attn.py, the container-free runner used for all measurements here, and the KLD harness are in the companion repo listed below.

What that command reproduces, and what it does not. It reproduces the recipe — composition, widths and byte budget — and not the checkpoint. The published bytes are the artifact. A fresh conversion of the hydrated sibling's recipe, run on the same hardware with the same flags and source, returned 13 of 16 pinned payload files identical (every config, the tokenizer, the index and both quantization descriptors) with byte-identical shard headers, the same tensor names, dtypes, shapes and offsets, and the same per-role byte totals and assigned widths — while 399 of the 409 quantized modules (97.6 %) differed inside their .trellis payloads, at 41-92 % of the bytes each (mean 82 %). No scale, norm, embedding, vision or BF16 companion tensor moved. The converter is nondeterministic, measured rather than inferred: two runs of one conversion, minutes apart, agreed on every width and every global scale and disagreed on the converter's own proxy_err (receipts/converter-determinism.json). So what you get from the commands above is a sibling: a different valid artifact of the same recipe, not a broken one. Every fidelity number on this card measures the published bytes a downloader receives and is unaffected. And "rebuild this and you get these numbers" is no longer an untested expectation: a third conversion of that hydrated recipe — a sibling again, differing in 399 .trellis payloads and nothing else — was captured and replayed on the identical v5 shard-0 protocol (512 contexts, 1,048,064 scored positions, same shared head, same comparator), and paired against the published checkpoint the difference is −3.755e-06, 95 % source-cluster bootstrap interval [−2.854e-05, +2.062e-05], which brackets zero, on 257 contexts to 255 with no ties. Two controls make that attributable to the sibling's weights rather than to the harness — replaying the published checkpoint against the same reference capture returned its mean and all 512 per-context rows bitwise, and a fresh recapture reproduces that mean exactly — so the comparison's floor is zero rather than a tolerance. 97.6 % of the quantized modules come back with different bytes and the fidelity is the same to within our resolution, which is what makes "the recipe is the reproducible thing, the bytes are the artifact" a measured claim rather than a hedge. It is one sibling, one recipe, one shard, at this protocol's resolution: it bounds the converter's fidelity variance, it does not estimate it, and it is not a finding that converter nondeterminism is fidelity-neutral in general (receipts/sibling-rebuild-fidelity.json). Check the digest of the published tree against SHA256SUMS rather than against your own conversion, and read a byte diff below this floor as the converter rather than as tampering, corruption or a changed recipe. The same run also showed the recorded build environment is incomplete: the pinned image has no marisa_trie, which the conversion imports on an unconditional path, so that image provably could not have finished the job. The conversion-capable image is gg-r34-convert. This is the converter's sense of reconstruction only — the runtime's reconstruct_had_slice / reconstruct_fp8_slice path, which turns stored trellis bytes back into weights at load, is a different claim and is untouched here.

Reproduce this

This section is about the numbers, not the bytes: a fresh conversion of the recipe produces a sibling rather than this checkpoint, as recorded under Reproducing this quant.

Everything the v5 numbers above were computed from is published as a dataset: malaiwah/qwen38-27b-fidelity-suite-v5 — 5,835 files, 10,826,796,868 B (10.83 GB / 10.083 GiB), verified at revision 08bde6cc. It contains the 5,120 token-id files that are the authoritative evaluation input (retokenizing the source text does not reproduce them), the parent suite manifest whose sha256 equals the ladder pin, the ladder pin itself, all ten 512-context shard views with their capture and replay command lines, the corpus fetch log, the shard-0 unquantized BF16 hidden-state reference (512 captures plus manifest, 10.73 GB) and 79 per-shard reports — 50 ladder, 10 tail, 15 scored-window, 4 cross-engine. Receipt receipts/preserved-artifacts.json; 5,326 of the 5,835 files were re-downloaded and re-hashed end to end, and the 10 GB hidden-state tree was checked against the Hub's own LFS digests plus a three-file CDN spot check.

What it costs to replay. Because the suite and the shard-0 BF16 reference are both published, a third party can score a new candidate against the identical contexts without recapturing the reference: one candidate capture plus one replay, about 6 minutes of GPU for shard 0 on a single RTX PRO 6000 Blackwell, instead of two model loads. That is exactly how the NVFP4 row above was produced. Re-running the whole ten-shard ladder is a different bill — about 5 hours of GPU for the fifty candidate captures, plus about 54 minutes for the nine BF16 shard references that were deleted once their reports verified.

Seven archival mirrors keep the third-party citations resolvable — the three explained below, plus the four quant-landscape shortlist mirrors cited elsewhere on this card, each pinned to the exact upstream revision measured and each now carrying a provenance banner that names its upstream author. malaiwah/Qwen3.8-27B-NVFP4-archival-9c73e2da is a recovery mirror: upstream super-squashed its history on 2026-08-15 and the Hub now answers Invalid rev id for 9c73e2da…, the revision every NVFP4 number on this card was measured against, so the reviewed revision is otherwise unreachable. malaiwah/Qwen3.8-27B-GGUF-archival-f1bfb127 is precautionary: the five files the cross-engine table cites, at a revision that still resolves upstream. malaiwah/Qwen3.8-27B-NVFP4-RTX5090-archival-69274a0d is likewise precautionary: the gittensor checkpoint measured above, at a revision that still resolves upstream, deep-verified after upload — its 19.2 GB of weights cost essentially zero transfer because the Hub already held every chunk. Said plainly, a mirror preserves the citation — a resolvable repo id, revision and digest table that survive an upstream squash or delete — and is not independent byte-level redundancy. Hub storage is content-addressed, so our copy and upstream's plausibly reference the same underlying chunks; nobody should assume physical copies we do not hold. The measured cost of the first two mirrors was 2.34 GB of transfer for 149.3 GB of content, about 1.6 %, which is that content-addressing showing through.

Prior art and credits

• turboderp / turboderp-org/exllamav3 — the EXL3 Trellis format, LDLQ calibration, MCG codebook, and the conversion pipeline.
• Martin Vit / voipmonitor / festr2 (local-inference-lab) — the Gilded Gnosis vLLM fork, B12X Trellis kernel integration, the voipmonitor/vllm Docker images, the rtx6kpro serving wiki, and the exl3-b6 online-encoding overlay whose distribution-fidelity protocol this evaluation follows.
• Luke Alonso / lukealonso — B12X kernel co-author (PR #31, NF3/NVFP4) and the NVFP4 quantization methodology whose protection patterns informed the recipe design.
• Brandon Music / brandonmusic — the ShapleyMCG expert-bit allocation method, the EXL3 checkpoint patterns (TR3 schema, online K6, shared-expert handling) that this recipe family is modelled on, and the exllamav3 SM120 retile fork.
• Josh Cartu / jcartu — the MTP78 recipe and rank-sliced MTP runtime that informed malaiwah/GLM-5.2-EXL3-TR3-MTP78: the graft/encode/serve procedure for the MTP draft layer, plus the vLLM normalize_rank_sliced_weight_name() loader patch.
• Qwen — the base model (Qwen/Qwen3.8-27B, Apache-2.0) and its official FP8 derivative (Qwen/Qwen3.8-27B-FP8).
• nvidia/Qwen3.6-27B-NVFP4 — the role-aware recipe this design is modelled on (4-bit MLP, 8-bit attention, protected embeddings/vision/MTP).
• unsloth/Qwen3.8-27B-NVFP4 — independent confirmation of the same protection pattern.
• malaiwah / progressive-tensors — research, recipe, harness and receipts. Per-expert EXL3 provenance work, the per-bit error ladder, the GLM-5.2-EXL3-TR3-MTP78 checkpoint, and the RTX 5090 context-capacity qualification that caught a 262K overclaim.

Companion repository

Recipe derivations, the measured composition of both NVFP4 references, the runtime contract, the toolchain gaps, the KLD protocol and the iteration log: https://github.com/malaiwah/qwen38-27b-exl3.

Successor checkpoints are published as separate repositories, each with its own measurement receipts (build-receipt.json, SHA256SUMS, quantization_manifest.json).