gguf-faultscope: single-bit fault exposure of block-quantized language models
Year: 2026
Type: software
Abstract: A tool and a dataset for the question of what a single flipped bit does to a block-quantized language model. Silent data corruption research measures fault effects on models held in 32-bit or 16-bit floating point, because that is how models are trained; deployed inference uses block quantization, where 32 or 256 weights share one 16-bit scale, and that sharing changes what one bit can reach. The tool derives a file's corruption exposure from its header alone, using the block layouts the runtime publishes, reports any tensor type it cannot model rather than silently excluding it, and checks a file's scales against a range recorded at build time. The harness injects single-bit faults stratified by structural role and by block type, restores and verifies each bit, and scores one forward pass per injection. In 6,725 injections across three campaigns on a Tesla P100, all 955 hard failures landed in the exponent field of a scale and 3,925 injections elsewhere produced none, including 2,150 into the sign and mantissa bits of the same scales, which reach the same weights. How far a bit reaches is not what decides whether it does damage. Within the exponent, which bits matter moves with model size: a sixteenfold change in one block's scale is nearly harmless at 0.5 billion parameters, at 2.1 percent, and breaks the 1.5 billion parameter model 22.8 percent of the time, on non-overlapping intervals. The quantization format a file is built in changes its exposure 2.6-fold at 0.5 billion parameters and 13.5-fold at 1.5 billion, because a larger model actually receives the format it asked for rather than a fallback. Because a scale is a small number by construction the damaging corruption is one-directional, so a five-bit maximum recorded in the header and compared at load catches 68.7 percent of hard failures and every one of 310 flips of the most significant exponent bit, with no redundancy. The repository is built to be added to: the record schema, the plan audit and the device table exist so a run from other hardware is a pull request.
Author: Jacob, Manu Nicholas
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204647700