implementation detail · filed under optimization
MXFP8
The MXFP8 format is used for selected layers, including Mamba output projections, to preserve information or support training stability.
Also called keep these layers in MXFP8.
- sources
- 2
- models
- 2
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 2
Documented in
Evidence
2 spans quoted from the sources, strongest treatment first.
select layers (including latent projections, MTP layers, QKV/attention projections, and embeddings) are maintained in BF16 or MXFP8 for training stability
usedmodel architecturein Nemotron 3 UltraNVIDIA
To prevent loss of information, we keep these layers in MXFP8.
usedoptimizationin Nemotron 3 familyNVIDIA
Filed alongside
Other methods under optimization :: training precision.
NVFP4BF16NVFP4 pre-trainingFP8 mixed-precision trainingFP4+FP8 mixed precisionFP8-precision reinforcement learningBF16 gradient reductionBF16 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