Training
Rotation / transform quant
QuantizationReferenceSuppress outliers via rotations (QuaRot, SpinQuant, QuIP#, FlatQuant) to enable low-bit activations.
- Relieves
- CapacityBandwidthCompute
- Targets
- Weights · Activations
- Format
- INT4 / W4A4
- Granularity
- Per-group + Hadamard / learned rotation
- Lifecycle
- PTQ · no gradients
- Calibration
- Small calibration set
- Compression
- ~4× (down to W4A4)
- Quality
- Recovers most FP16 at 4-bit incl. activations
- Hardware
- INT4/FP4 kernels + Hadamard transform support
- Runtimes
- vLLMTensorRT-LLM
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.