implementation detail · filed under optimization
Quantize-Dequantize Training
Applying quantize-dequantize to experts after each parameter update to maintain training-inference consistency.
Also called quantize–dequantize (QDQ).
- source
- 1
- models
- 2
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
core 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
we apply quantize–dequantize (QDQ) to the experts after each parameter update.
coresoftware implementationin MiMo-V2.6Xiaomi
Filed alongside
Other methods under optimization :: quantization-aware training.
Quantization-Aware TrainingFP4 QuantizationFP4 Quantization-Aware TrainingMXFP4 Weights with MXFP8 ActivationsStochastic RoundingINT4 Quantization-Aware TrainingMXFP4 Quantization-Aware Post-TrainingNVFP4 TrainingPer-Block Scalar ScalingQ4_0 Quantization FormatRandom Hadamard TransformsStochastic Rounding for Mamba CacheStochastic Rounding of GradientsStraight-Through Estimator