Alberto-Codes/gemma-4-31B-it-fit24gib-GGUF

<!-- Authored for issue #446 under the #401 identity grammar. Upload this file verbatim — the published card and this source must match. The number ledger lives beside this file in the vramfit repository: publication/gemma-4-31b-fit24gib/card-ledger.md. -->

gemma-4-31B-it-fit24gib-GGUF

This repository carries a mixed-precision GGUF of google/gemma-4-31B-it, packed from the QAT unquantized checkpoint and solved to serve images inside a 24 GiB card. vramfit measured each decoder layer's quantization damage — the shift in the model's output distribution when that layer quantizes — then allocated bits under the budget and packed through llama.cpp's quantizer. The fit24gib marker is a deployment claim, not a file size: a serve test ran these exact files inside a 24 GiB VRAM boundary under the configuration stated below. This pack has no single quantization scheme. The budget is the claim.

Two files, one artifact. The decoder (gemma-4-31B-it-fit24gib.gguf, 14.92 GiB) carries the quantized language model. The projector sidecar (gemma-4-31B-it-mmproj-q4km.gguf, 629 MiB) carries the vision tower, converted from the vendor projector with llama-quantize's Q4KM recipe. The recipe name labels the command, not the contents. Every quantizable tensor falls back on this geometry. The file holds 150 Q50, 13 Q80, 27 F16, and 166 F32 tensors at 9.16 effective bits per parameter, read from its own header. Text-only serving needs the decoder alone. Image serving needs both files. The sidecar shipped in BF16 until 2026-08-31, when a measurement priced the trade. The converted sidecar matches the BF16 sidecar on content-class KLD (0.0050 against 0.0045, 99.2 % top-token agreement both). It buys 4,096 tokens of context at the 2026-08-31 serve boundary. The record is ADR-0030 open question 2 and its 2026-08-31 amendment.

The headline

The official QAT Q4_0 GGUF of this model is 16.44 GiB. This pack is 14.92 GiB at comparable measured text quality — four benchmark ties and one win, tables below — and lower measured vision-conditioned divergence. The freed weights buy context on the same card:

Both rows are served results on an RTX 4090 (24 GiB), llama.cpp b10362, -ngl 99 -np 1, measured 2026-08-31 at 4,096-token ladder rungs in one frame — one day, one idle VRAM share, both packs. The QAT image row serves the vendor's own BF16 projector, so that row moves two variables: the decoder and the sidecar. The same-sidecar comparison holds one variable. Behind the same BF16 projector this pack serves one image at 69,632 against QAT's 49,152, a gain of 20,480 tokens (+41.7 %). The sidecar conversion adds the last 4,096. Every measured number below sits beside its baseline counterpart. The card prints the losing numbers too.

Quality beside size — text

Measured on the framed calibration corpus: 357 blocks of public- domain text under the Gemma channel frame, 182,404 tokens, 356 chunks at nctx 512. This is the pack's measurement frame — the same corpus the importance matrix consumed — not held-out text. The held-out check is the benchmark table below. The instrument is a CPU-only llama-perplexity built from the b10362 source with `parsespecial` enabled, and it measured both sides of every ratio. All three rows share the KLD-pass perplexity convention. Lower is better for PPL and KLD. "Same top" is the share of tokens where the pack and the BF16 reference agree on the top token.

Read the split honestly: this pack holds the better PPL ratio, and the QAT baseline holds a slightly better mean KLD (0.0420 against 0.0446) and top-token agreement (92.32 % against 92.04 %). One asymmetry runs in this pack's favor: its importance matrix consumed this same corpus, and the QAT baseline consumed no matrix from this frame, so tier 1 and tier 2 lean toward this pack. The two text metrics disagree at the margin, and the held-out benchmarks arbitrate.

Held-out benchmarks

lm-evaluation-harness 0.4.12 over llama-cpp-python 0.3.34 with the b10362 Vulkan libraries, full splits, no sampling limit. A delta inside the combined standard error is a tie. The Δ column computes from the sidecars' unrounded scores.

Four ties and one win, at 1.5 GiB fewer packed bytes. The two nominal deficits — HellaSwag −0.63 and GSM8K −0.08 — sit inside their combined standard errors. Per-task records ship in the evals sidecars beside the weights.

The measured vision bound

This card states a vision-quality bound because a campaign measured one for this exact artifact. The sensitivity map that solved this recipe measured text damage only — a text-measured map licenses no vision-quality claim — so the bound below comes from its own image-conditioned measurement, not from the map.

The metric is a truncated top-20 KLD in nats at the serve boundary. The llama.cpp server caps n_probs at 20, so this is not a full-vocabulary KLD. A BF16 reference decoder generated greedily over 10 held-out 768×768 images, and the harness teacher-forced each quantized arm on the reference sequence. The campaign's three decoder arms served through the same BF16 projector. A fourth arm measured the shipped pair — this decoder behind the Q4KM sidecar — with the projector as the only variable. Position classes separate image-grounded content tokens from channel-frame policy tokens (markup, pos0). Frame-policy positions dominate all-position means on this channel-locked model, so the content class carries the vision claim.

Content-class results (n = 120 positions):

The shipped arm measures 0.0050 against the BF16-sidecar arm's 0.0045 at the content mean, and 0.0193 against 0.0239 at p95. The all-position means sit at 0.0483 (shipped) and 0.0489 (BF16 sidecar) — 9.7 and 10.9 times their content-class figures, with frame-policy positions dominating both. The QAT baseline's all-position mean over the same 178 positions is 1.1092. The instrument noise floor is 1.07e-4 mean KLD, measured by teacher-forcing the BF16-sidecar arm on its own greedy sequence. The shipped arm measures 7.5 times below the QAT baseline and 47 times above the floor. The BF16-sidecar arm measures 8.3 and 42. The BF16 reference and the BF16-sidecar arm each answered 10 of 10 held-out image questions correctly on their own generation paths.

Two caveats travel with these numbers:

• —The BF16 reference ran on CPU — a second instrument. The KLD figures read as divergence from the reference distribution, never as same-instrument damage numbers. Only the pack-against-QAT comparison is same-instrument.
• —The bound belongs to this artifact and this metric. A different recipe on the same checkpoint measures its own bound or states none. No measurement covers transfer between artifacts.

The full campaign record — per-image tables, position-class breakdown, the 178 per-position KLD pairs, input log hashes — ships in analysis/vision-campaign-kv9.json. The pairs recompute every derived number for the BF16-sidecar arms. The shipped arm's record lives in the vramfit tracker (#451) and in ADR-0030 open question 2, with raw logs in the run archive.

The real-GUI campaign

The vision bound above rests on 10 held-out 768×768 images. A second campaign measured three arms on real GUI screenshots at 1280×720. The dataset is PSAI Computer Use Data (anaisleila/computer-use-data-psai, MIT): the 1,349 tasks that carry both screenshots and interaction events. The set is 1,318 EASY tasks and 1,346 browser tasks out of 1,349. The MEDIUM and HARD rows exist in the artifact at 15 and 16 tasks, too few to read, and do not print here.

One H100 SXM 80 GB ran llama.cpp b10362 CUDA at -ngl 99, context 8,192, -np 1, greedy, with a 48-token generation cap. All three arms (the BF16 reference, this pack, and the QAT Q4_0 baseline) served through the same BF16 projector on that one instrument. The vision bound above could not claim that: its reference ran on CPU.

Divergence. The metric is the vision bound's: a teacher-forced truncated top-20 KLD in nats at the serve boundary, 27,962 positions, with the same position classes. The content class carries the claim.

Read the two scales apart. On content positions the QAT baseline's mean is 1.18 times this pack's, and its median 3.6 times, from the artifact's unrounded values. This campaign measured no instrument noise floor. The prompt-prefix difference disclosed below rides inside the mean margin. Read the content-class gap as a bound, not as a separation. The all-position gap of 4.5 times is a frame-policy number. The QAT baseline diverges on the channel-frame positions (markup mean 1.13, pos0 mean 1.61), where this pack measures 0.03 and 0.11. The vision bound's 0.0045 content mean comes from another frame and another instrument. The two figures do not compare.

Task identification. A judge scored each arm's one-sentence answer against the task's ground-truth name. A MATCH needs the same application or site and the same user goal. Over 1,349 tasks the BF16 reference scored 49.9 %, this pack 51.2 %, and the QAT baseline 51.4 %. The QAT scores come from a 512-token regeneration. All 1,349 of its 48-token generations ended with no answer. No arm separates. The judge's noise floor is 6.9 %. On 598 tasks the reference and this pack wrote byte-identical answers. The judge returned different verdicts on 41 of them. Every gap between the three scores sits inside that floor. The ceiling is the frame — one middle screenshot against a full-task name left the BF16 reference itself at 49.9 %. The content-class divergence gap does not surface as an accuracy gap in this frame.

Throughput. Per-request server timings, means over generations, one slot, both quantized arms behind the BF16 projector:

The H100 timings are incidental, not a controlled timing arm, and the H100 is not the target card. The 4090 arms ran the same harness, prompt, generation cap, and 20-task subset. That card is this pack's target: b10362 Vulkan, -ngl 99 -np 1, context 8,192, under desktop sharing. No BF16 reference arm ran on the 4090. The gap inverts. This pack decodes 10 % faster than the baseline on the 4090, where the H100 CUDA build had the baseline 15 % ahead. The spread over 20 generations is 0.5 tok/s standard deviation (pack) and 0.2 tok/s (QAT). The shipped sidecar measures 47.7 tok/s, inside that spread. One asymmetry runs against this pack. Its generations stop at a mean 20 tokens, and the QAT generations hit the 48-token cap. The QAT arm therefore amortizes per-request overhead over more tokens.

Five disclosures ride every number in this section:

• —Screenshot policy. One middle screenshot per task, 1280×720. One such screenshot is 271 decoder tokens on this pack, measured at the server. The vision bound's images are 768×768.
• —Reference origin. convert_hf_to_gguf.py produced the BF16 reference decoder from the QAT unquantized checkpoint at b10362. The H100 reference and the vision bound's CPU reference are two instruments.
• —Template confound. The three GGUFs render one identical request to 319 (reference), 324 (this pack), and 312 (QAT) prompt tokens, uniform across all 1,349 tasks. Each arm's KLD folds that prefix difference into the measured divergence.
• —Generation budget. Every QAT 48-token generation ended inside the thought channel, the model's reasoning block before its answer. The 512-token regeneration ran on a second H100 SXM pod of the same SKU and build. 49 of its 1,349 generations still hit the cap, and the judge scored those as fragments. The reference and this pack answered under 48 tokens on every task.
• —Judge transport. The judge is claude-haiku-4-5 through the Claude Code CLI (claude -p), which exposes no temperature control. All 4,047 verdicts persist in the run archive.

The full record ships in analysis/psai-gui-kv9.json: method, 13 input log hashes, aggregates, and 1,349 per-task rows with every forced position. The per-task rows recompute the divergence tables and the accuracy scores. The noise floor traces to the answer files in the run archive, with the raw generations, force logs, and judge outputs. The tracker record is #462.

What fit24gib means

The claim: the decoder loads fully offloaded on a 24 GiB card at 86,016 tokens of context and generates, and the decoder plus the projector load at 73,728 tokens and answer an image prompt. Both are measured serve results under the stated configuration, not a promise about every runtime setup.

The budget arithmetic behind the recipe:

The packed file lands 86.08 MiB under the weight budget.

The serve ladders, llama.cpp b10362 Vulkan on an RTX 4090 (24,564 MiB) under desktop sharing, -ngl 99 -np 1, KV cache f16, 4,096-token rungs, measured 2026-08-31 with 23,629–23,631 MiB free before each load:

The ladder boundary moves with the box's idle VRAM share. The 2026-08-28 card published 81,920 text-only with 86,016 as its failing rung, and 61,440 with one image behind the BF16 sidecar. Both boundaries are real in their frames. The published 81,920 rung reproduced on 2026-08-31 (465 MiB free at load) before these ladders ran, and 86,016 then passed with 143 MiB free. The text boundary moved one rung with the frame. This pack's image boundary moved three rungs: one from the frame, one from the BF16 line re-measuring lower in this frame, and one from the sidecar conversion. At the 86,016-token boundary the decoder answered a completion request from inside that envelope. The serve ladder is a fit bar, not a speed bar — the boundary check decoded five tokens under desktop sharing. The throughput figures sit in the real-GUI campaign section above, measured at context 8,192. Throughput at the 86,016-token boundary is unmeasured.

Serving images costs a measured 960 MiB beyond text-only serving with the shipped sidecar: a 772 MiB load-time delta at matching context, plus the image-encode transient. The line quantizes to the 4,096-token rung, 320 MiB on this geometry. The BF16 sidecar's line measured 1,280 MiB in the same frame. Its components measured 2026-08-28: 1,022.8 MiB of projector weights and a 150.63 MiB CLIP compute reserve, plus the transient. One 768×768 image consumes 256 decoder tokens, measured at the server — the checkpoint config claims 280, and the measured cost wins.

To serve:

# Text only, at the measured boundary
llama-server -m gemma-4-31B-it-fit24gib.gguf -c 86016 -ngl 99 -np 1

# Images, at the measured boundary
llama-server -m gemma-4-31B-it-fit24gib.gguf \
  --mmproj gemma-4-31B-it-mmproj-q4km.gguf -c 73728 -ngl 99 -np 1 \
  --mtmd-batch-max-tokens 264

Three reproduction traps, stated because each cost a failed load or a crashed server:

• —Pass -np 1. The b10362 server defaults to -np 4 with unified KV, which adds ~2,400 MiB of SWA cache for this geometry and fails loads that fit at one slot.
• —Keep ~200 MiB free beyond the load when serving images. The image-encode transient allocates at request time, and the b10362 server crashes on the failure path instead of refusing.
• —Cap the encode batch at one image with --mtmd-batch-max-tokens 264. A 1280×720 image is 264 image tokens, 271 with its wrapper. The b10362 server packs up to 1,024 image tokens into one encode graph by default. Two such images then share one graph, and its compute buffer asks 328 MiB against the 150.63 MiB one-image reserve. At this boundary the server holds ~100 MiB free after one image on this pack and ~65 MiB on the QAT baseline. A 2026-09-02 multi-image ladder (23,549–23,556 MiB free before each load) crashed the server on the second image at both configurations in the table above. With one image per batch the same ladder filled the window to the context refusal on both, and the server survived every request. The ~200 MiB rule above is a one-image rule.

The recipe

The solver — greedy damage-per-byte over the sensitivity map — allocated 61 decoder groups: the token embedding and 60 layers. A nominal width names the solver's assignment, and the packer maps it to a k-quant type under the recipe's kquant-imx within-group policy. The allocation:

• —6 layers at nominal 2 (Q2_K): layers 1–5 and 11.
• —9 layers at nominal 3 (Q3_K): layers 0, 6–10, and 12–14.
• —45 layers at nominal 4 (Q4_K): layers 15–59.
• —The token embedding at nominal 4 (Q4K), with the output head at Q4K by the same override.

The map priced the early layers cheapest, and the solve spent its budget protecting depth. The plan's trace records 81 steps and replays from recipe.json, which ships in this repository. The predicted whole-recipe damage is 0.184 on the map's frame. Damage values are one scan's measurements on one measurement frame — do not compare them across scans or across models.

The pack ran llama-quantize b10362 with --pure, a Q2_K base type, and 60 per-tensor overrides from the recipe.

The importance matrix

The pack consumed an in-frame importance matrix: 356 chunks of the framed calibration corpus through the pinned prebuilt llama-imatrix b10362 with --parse-special. The matrix stays in the run archive and is not carried in this repository.

Coverage derives from the matrix's own entry names, never from a label. Two absences matter:

• —The matrix carries no token_embd.weight entry, so the token embedding quantized unassisted. This is expected at b10362.
• —The matrix carries attn_v entries for 50 of 60 layers. The 10 fullattention layers (5, 11, 17, 23, 29, 35, 41, 47, 53, 59) receive no `attnv` activation from the b10362 graph, so those ten tensors quantized unassisted. This is a property of the instrument, not a defect in the matrix.

Reproduce it

The repository ships the recipe, the run log, the evals sidecars, and the two analysis artifacts beside the weights. Two commands rebuild the decoder from the base checkpoint:

python convert_hf_to_gguf.py <checkpoint dir> \
  --outfile gemma-4-31b-it-qat-unquantized-bf16.gguf --outtype bf16
uv run vramfit pack recipe.json --llama-cpp <llama.cpp checkout> \
  --model <checkpoint dir> \
  --base-gguf gemma-4-31b-it-qat-unquantized-bf16.gguf \
  --imatrix gemma-4-31b-bf16-framed.imatrix.gguf \
  --out gemma-4-31B-it-fit24gib.gguf

The recipe records the full solve: the 15 GiB weight budget, the 0.005 format overhead, and the 81-step trace, so the type placement replays from recipe.json alone. Two inputs stay in the project's run archive and are not published: the importance matrix and the sensitivity map. Without the matrix the pack step reproduces the type placement but not this file's exact bytes, and without the map vramfit plan cannot re-derive the recipe or solve a different budget for this model.

Damage disclosure

Quantization compresses every weight tensor with one uniform lossy procedure. It does not bypass or disable the base model's safety training, which ships in these weights at lower precision — and it can shift any model behavior. The tables above are the measured bound on that shift. Tier 2 measures whole-model KL divergence against the BF16 reference on the measurement frame. Tier 3 holds four ties and one win on held-out benchmarks. The vision campaign bounds the distributional shift on 120 content-class positions over 10 held-out images. The real-GUI campaign bounds it on 21,216 content-class positions over 1,349 screenshots.

One limit, stated plainly: damage measures output distributions, not safety behavior separately, and the vision bounds are distributional bounds on images, not a safety evaluation of image inputs. Read this card as a damage disclosure, not a safety certificate. Deploy this pack with the same system-prompt and application-layer protections you would give the base model.

Files

The run log ends at pack_finished: this pack predates the vramfit release that records the sidecar hash in the run log, so the projector hash above was computed directly on the shipped bytes. Until 2026-08-31 this repository shipped the vendor's BF16 projector (gemma-4-31B-it-mmproj.gguf, SHA-256 6bd60bdb…07a4, 1,200,726,368 bytes). ADR-0030's 2026-08-31 amendment swapped the sidecar to Q4KM. The vendor repository carries the identical BF16 bytes.

Provenance and license

The source checkpoint is google/gemma-4-31B-it-qat-q4_0-unquantized at revision 1e4d8beecacb8b7590c1d8bedd7335f687bf311f. Conversion to the BF16 decoder GGUF ran at llama.cpp b10362. The projector derives from the vendor's own published file, downloaded from google/gemma-4-31B-it-qat-q4_0-gguf at revision 59dde24573e7e61570dba08b18a2e1fe246955ed. The source file's SHA-256 matches the vendor's LFS object. It carries 190 BF16 and 166 F32 tensors under the v. and mm. roots. llama-quantize b10362 converted it with the Q4KM recipe, 1,145.08 MiB to 628.96 MiB, with every quantizable tensor on fallback. The shipped type counts appear in the two-files section above. The QAT Q4_0 comparator is the vendor's decoder GGUF from the same repository at the same revision, SHA-256 179cfb99212709597eae5929112cfca677e1bbf566178b479ae1da0c4772874b.

Gemma 4 is released by Google DeepMind under the Apache 2.0 license — see the Gemma 4 license note and the LICENSE file in this repository. This derivative work carries the same license. The calibration corpus is public-domain Project Gutenberg text under the Gemma channel frame.