implementation detail · filed under optimization
BF16 gradient reduction
Gradient reduction or local accumulation uses BF16 precision rather than FP32.
- source
- 1
- model
- 1
- labs adopt it
- 0
- strongest
- not used
How sources treat it
One count per evidence span, weakest treatment to strongest.
not used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
The first divergence, which occurred at around 8T tokens, was attributed to a reduction in local gradient accumulation precision for the output layer from FP32 to BF16
not usedoptimizationin Nemotron 3 UltraNVIDIA
Filed alongside
Other methods under optimization :: training precision.
NVFP4BF16NVFP4 pre-trainingFP8 mixed-precision trainingFP4+FP8 mixed precisionFP8-precision reinforcement learningMXFP8BF16 mixed-precision trainingBlock-wise FP8 activation quantization with offloadE2M1FP32 attention-output retentionFP32 gradient reductionFP8 storage for the residual stateHigh-precision final network layersMixed-FP8 quantizationMixed-precision trainingNVFP4 fine-grained micro-block scalingTwo-dimensional block quantization