specific method · filed under optimization
Stochastic Rounding
A rounding method used with low-precision values; the evidence includes gradients and FP16 Mamba-cache use but does not establish one narrower shared application.
- sources
- 2
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
optional 1used 2
Documented in
Further reading
Picked by hand, not extracted: where to read more, not evidence for anything on this page.
- Pretraining Large Language Models with NVFP4 paper arxiv.orgshows stochastic rounding on gradients is essential for NVFP4 training stability
Evidence
3 spans quoted from the sources, strongest treatment first.
stochastic rounding on gradients
usedoptimizationin Nemotron 3 UltraNVIDIA
FP16 with stochastic rounding preserves FP32-cache accuracy
usedpost trainingin Nemotron 3 UltraNVIDIA
--enable-mamba-cache-stochastic-rounding
optionalinference servingin Nemotron 3 UltraNVIDIA
Filed alongside
Other methods under optimization :: quantization-aware training.
Quantization-Aware TrainingFP4 QuantizationFP4 Quantization-Aware TrainingMXFP4 Weights with MXFP8 ActivationsINT4 Quantization-Aware TrainingMXFP4 Quantization-Aware Post-TrainingNVFP4 TrainingPer-Block Scalar ScalingQ4_0 Quantization FormatQuantize-Dequantize TrainingRandom Hadamard TransformsStochastic Rounding for Mamba CacheStochastic Rounding of GradientsStraight-Through Estimator