malaiwah/fruit-fidelity-quant-siq-v1

GLM-5.2-SIQ-Fruit — quantized fidelity dataset (hidden form, reconstructed weights)

The candidate half of a three-step fidelity measurement on malaiwah/GLM-5.2-SIQ-Fruit @ c1798e3676fa16b4a874381171adab1e3033fbd5, captured on the same panel, the same lane and the same engine as its root, malaiwah/fruit-fidelity-root-v1.

step 1  capture   reference weights + panel  ->  fruit-fidelity-root-v1
step 2  capture   quantized weights + panel  ->  THIS REPOSITORY
step 3  compare   root, this                 ->  KLD + determinism + a registry receipt

The number

mean tokenwise KLD, KL(reference ‖ candidate)	0.038737 nats
top-1 agreement	87.98 %
median / p95 / p99 / p99.9 / max	0.016764 / 0.147713 / 0.317493 / 0.756305 / 1.496166
scored positions	32,752 over 16 windows
per-window mean, sd	0.038737, 0.028309
literary windows / scientific windows	0.027502 / 0.049973
estimator	full 154,880-column vocabulary, fp64 accumulation, shared reference head
comparability	strict, same_lane: true
floor	exactly 0.0 — root vs a second cold capture of the same weights

Read the scatter before the mean. Sixteen windows have a per-window standard deviation of 0.0283 around a mean of 0.0387 — a standard error of about 0.0071, so this number is worth roughly ±0.014 at 95 %, and the scientific windows sit at nearly twice the literary ones. It is a real measurement of this artifact on this panel. It is not precise enough to rank two artifacts that differ by less than about 30 %, and nothing here should be read as a quality score.

What was actually quantized

Only one tensor class. Everything else in the artifact is bf16 and byte-shared with the reference export.

class	treatment	format	bpw	layers
moe.experts	quantized	exl3-trellis	3.375	3–12
mtp	quantized	exl3-trellis	3	13 (not executed on this lane)
everything else	native	bf16	16	all

3.375 bpw is arithmetic, not a claim: the artifact's own tier_bitmap.json allocates 96 experts at K4 and 160 at K3 in each of the ten sparse layers, and (96×4 + 160×3)/256 = 3.375.

The config declares a scheme it does not use

config.json carries quantization_config with quant_method: modelopt, quant_algo: NVFP4, group_size: 16, W4A4 and producer b300-exl3-modelopt-dispatch-shim. That block does not describe the stored bytes. Zero tensors are NVFP4; the routed experts are exl3-trellis atoms. The exporter copies the block from the parent GLM-5.2 configuration rather than authoring it — its ignore list still names model.layers.78.eh_proj, and this model has thirteen layers. The scope above describes the bytes.

Reconstructed, not executed — read this before citing the number

The artifact stores each routed expert projection as four tensors, .rank0.{trellis,suh,svh,mcg}. Stock transformers cannot read that, and it does not fail: it reports model.layers.{3..12}.mlp.experts.{gate_up,down}_proj as MISSING, randomly initialises them, and returns a model that runs. A capture of that would be a confident number about weights nobody measured.

So this capture ran a bf16 reconstruction of the stored atoms, decoded by k6/tools/materialize_exl3_experts.py. That is the same dequantize-and-run methodology the GGUF, MLX and EXL3 ecosystems use for KLD, and it isolates weight error from kernel error. It is not the vendor runtime: Fruit's production path is b12x/SparkInfer + vLLM with fp8/nvfp4 KV cache and MTP speculative decoding, none of which is exercised here.

Evidence that the decode is the right decode

check	result
codebook LUT vs the campaign's independently frozen table	bitwise equal, all 65,536 entries
bit rate read off each payload vs the producer's tier_bitmap.json	8,448 / 8,448 agree
our reconstruction error vs the producer's own recorded expert_rel_rt_mse	ratio mean 1.00013, range 0.98902 – 1.01337 over 8,448 matrices
K4 experts (2,880) vs the bf16 reference	cosine 0.99773, rel-L2 0.06738
K3 experts (5,568) vs the bf16 reference	cosine 0.99124, rel-L2 0.13205

The third row is the decisive one: expert_rel_rt_mse is the encoder's own per-expert record of its reconstruction error, written at quantization time by a program we did not write. Reproducing it to 0.013 % on average from the published bytes is not something a wrong codebook or a wrong unpack can do.

The codebook table itself is transcribed from exllamav3 v1.4.2 exllamav3/exllamav3_ext/quant/codebook.cuh, decode_3inst<1>, and is pinned in-tree by digest. It has not been proven bitwise against a running exllamav3 CUDA kernel on this hardware; the evidence above is what stands in for that, and this caveat is recorded in the dataset's disclosures.

The panel

panel--fruit.malaiwah.heldout-v1 — 16 windows × 2,048 tokens, built for Fruit and for nothing else, suite_token_hash_sha256 a6d367cc3ba448800372dee435d2bb4f536d23ca68843628832fa3b122ceabe1. Its build receipt ships inside this dataset at panel/panel-receipt.json. See the root dataset's card for how it was constructed and what its separation claim is.

Reproduce the comparison

Neither set of weights is needed:

python3 bin/fidelity_dataset.py compare \
    --reference hf://malaiwah/fruit-fidelity-root-v1 \
    --candidate hf://malaiwah/fruit-fidelity-quant-siq-v1 \
    --out cmp --verify-tensors

To rebuild the candidate from the published artifact:

python3 k6/tools/materialize_exl3_experts.py \
    --src <GLM-5.2-SIQ-Fruit checkout> --out siq-bf16 --device cuda \
    --reference <GLM-5.2-SIQ-Fruit-bf16 checkout> \
    --tier-bitmap <GLM-5.2-SIQ-Fruit>/tier_bitmap.json \
    --receipt siq-reconstruction-receipt.json

python3 bin/fidelity_dataset.py capture \
    --out ds-siq --form hidden --role quant --lane local-cuda-budget \
    --engine hf-transformers -- \
    --model siq-bf16 --weights-repository malaiwah/GLM-5.2-SIQ-Fruit \
    --model-revision c1798e3676fa16b4a874381171adab1e3033fbd5 \
    --panel <panel dir> --panel-role final --device cuda \
    --panel-id panel--fruit.malaiwah.heldout-v1 \
    --scope-file <scope-siq.json> --codec exl3-trellis --declared-bits 3.375 \
    --dataset-id fidelity--fruit.malaiwah.quant.siq-exl3-reconstructed \
    --dataset-name "GLM-5.2-SIQ-Fruit (exl3-trellis K3/K4 experts, bf16-reconstructed) fidelity dataset"

Verify what you downloaded

python3 bin/fidelity_dataset.py verify <this directory>

Disclosures

• reconstructed_weights (caveat, affects comparability) — see above.
• declared_scheme_mismatch (caveat) — the artifact's quantization_config describes NVFP4/modelopt; the bytes are exl3-trellis.
• architecture_subset_loaded (info) — stock transformers drops the DSA indexer for layers 3–13 and the whole MTP layer 13, identically for the reference and this candidate.
• reduced_run_count (caveat) — one cold capture.

MIT licensed, like the tooling.