NVIDIA-Nemotron-3.5-Lightning-30B-A3B-sensitivity-maps

This dataset carries the expert-stack quantization sensitivity maps of nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 (revision ce38b6a). vramfit measured them. A sensitivity map records one damage number per layer group and candidate precision. Damage is the shift in the model's output distribution when that group alone quantizes — mean final-logits KL divergence against the bf16 reference. The maps describe the base model, not any quantized file. They contain no model weights.

The packed model solved from the q0-imx map below ships as NVIDIA-Nemotron-3.5-Lightning-30B-A3B-fit16gib-GGUF.

The map format

Each map is one JSON file, vramfit_schema 3, in the form that carries group_by: stack. The scan block records the measurement frame: metric, calibration file, token count, candidate precisions, grouping, within-group method, and imatrix path. The groups list records the model's 46 routed-expert stacks. An expert stack is the 128 routed experts of one projection in one layer, stored as a single tensor — the unit a GGUF pack assigns a precision to. Each group carries its member tensors, its bytes at reference precision, its damage per precision, and tensor_bytes, the per-tensor size split. The q0-imx map also carries imatrix_counts. The sensitivity map format page specifies every field. The path fields record the measuring machines' absolute paths. The scan.calibration basename matches the file here. The scan.imatrix matrix is bartowski's, published in bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF.

The maps hold the expert stacks only. Dense groups never enter them: the published recipe pins every quantizable dense class at 8-bit and prices it from the checkpoint's safetensors headers.

The two scans

File	Calibration tokens	Within-group method	Imatrix	Started (UTC)	Cells
sensitivity-32k-q0-imx-stacks.json	32,768	q0-imx	yes	2026-08-21	92
sensitivity-32k-q0-ref-stacks.json	32,768	q0-ref	no	2026-08-18	92

Each scan covers the 46 expert stacks at candidate precisions {4, 2}, which is 92 cells. The q0 method round-trips each cell through llama.cpp's _0-family quantizers (one fp16 scale per block, no super-block), ported against the C reference and bit-exact for the unassisted types. q0-ref fits unassisted. q0-imx weights the nominal-4 fit with the pack's importance matrix. The 2-bit quantizer ignores the matrix, so the two methods differ at nominal 4 only. bartowski publishes the importance matrix the q0-imx map and the pack consumed, and the model card pins its SHA-256.

The two maps play different roles in the publication:

• The published recipe solved from the q0-imx map. recipe.json in the model repo records the full solve: the budget bytes, the nine pins, the format overhead, and the 11-step trace.
• The q0-ref map is the attribution bound. Its damage ordering agrees through every rank the solve reads, and it derives the identical placement. So the published win credits the damage ranking under the placement rule, not the importance matrix.

Do not compare damage across files

Damage values are calibration-relative and frame-relative. They compare only within one file. Do not rank damage across scans, across calibration sets, or across models. The two scans exist to separate the ranking from the imatrix, not to be averaged. Rank packed models by measured quality at a fixed model and budget, never by raw damage.

Solve a recipe

vramfit plan is pure Python and imports no torch. Solve your own budget against the q0-imx map:

uv run vramfit plan sensitivity-32k-q0-imx-stacks.json \
  --checkpoint <checkpoint dir> --vram 16GiB

--checkpoint prices the dense groups the map does not hold. An unstated --kv-headroom defaults to 4GiB, so this example solves a smaller weight budget than the publication. The published solve passed the ruled 15.776 GiB weight budget as --vram with --kv-headroom 0, nine dense pins at 8-bit, and a 0.002 format overhead. recipe.json in the model repo records every resolved value.

Run logs

Each scan ships its run log, <scan name>.runlog.jsonl — structured JSONL, one cell_measured event per cell between the lifecycle events scan_started, meter_built, and scan_finished. Every line carries vramfit_runlog 2. Both scans ran start to finish with no halt: 92 cell events each. Every cell event records the group, the bits, the measured damage, the wall-clock seconds, and the process memory high-water mark.

The calibration set

calibration.txt is the complete Project Gutenberg ebook of Pride and Prejudice, unmodified, with the Project Gutenberg header and license text intact — byte-identical to the calibration file of the project's first publication. Both scans name this file in scan.calibration and read 32,768 tokens of it. The evaluation tiers ran the packed model on held-out WikiText-2 test text, never on this file.

Files and hashes

File	SHA-256
sensitivity-32k-q0-imx-stacks.json	68c79d24b87dd58012ac72e0a1c6d0137173b10b528057afee36bf2079bd00e5
sensitivity-32k-q0-imx-stacks.runlog.jsonl	8f74d51303c66c9b8d607ee53122a137fa7c9707495b2c9edc19c8d63b8dc60a
sensitivity-32k-q0-ref-stacks.json	0fbcf84fc5d471fbc7a943485256ac8521596cf605b3c836bc1622e6c6b4f55e
sensitivity-32k-q0-ref-stacks.runlog.jsonl	1ce8fd3d4287abb240a314f3972fbf756279f74feb9eadd934b1aa126992b073
calibration.txt	74f2665d6e6925fc2c17dec644bec9e87df478a0f1836822125e8acbb3777806

Hashes prove identity, not quality. The measurement evidence is the run logs beside each map.

License

The maps and the run logs are CC-BY-4.0. calibration.txt is a Project Gutenberg ebook, public domain in the United States, distributed with its Project Gutenberg header intact. This dataset carries measurements of the base model, not the base model's weights. OpenMDW 1.1 governs the base model repo and the packed-model repo.

Disagree with a number?

Re-run the scan. The vramfit repository documents the scan command, the meter, and the settings the run logs record. A map you measure yourself beats one you argue with.