howtospark
Training

SparseGPT

Pruning / sparsityReference

One-shot layerwise unstructured pruning to 50% with error compensation.

Relieves
Capacity
Targets
Weights
Format
Unstructured sparsity
Granularity
Unstructured (≈50%)
Structure
unstructured
Lifecycle
PTQ · no gradients
Calibration
Small calibration set
Compression
~2× (50% sparse)
Quality
Minimal loss at 50%; stacks before quant
Hardware
Sparse kernels for speedup, else capacity-only
Runtimes
Transformers

A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.