Techniquesoptimizationquantization-aware training

implementation detail · filed under optimization

Per-Block Scalar Scaling

Introducing a scalar scale at each block to bound activation ranges for stable FP16 inference.

source: 1
model: 1
lab adopt it: 1
strongest: used

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 1

Documented in

Evidence

1 span quoted from the sources, strongest treatment first.

to enable stable inference in fp16, we introduce a scalar scale at each block in order to bound the activation ranges to fit fp16.

usedinference servingin Gemma 4Google DeepMind

Gemma 4 Technical Report·Gemma Team, Google DeepMind·2026-07-02

Filed alongside

Other methods under optimization :: quantization-aware training.

Quantization-Aware Training FP4 Quantization FP4 Quantization-Aware Training MXFP4 Weights with MXFP8 Activations Stochastic Rounding INT4 Quantization-Aware Training MXFP4 Quantization-Aware Post-Training NVFP4 Training Q4_0 Quantization Format Quantize-Dequantize Training Random Hadamard Transforms Stochastic Rounding for Mamba Cache Stochastic Rounding of Gradients Straight-Through Estimator