specific method · filed under optimization
FP8 mixed-precision training
Training uses mixed precision with FP8, without further precision-specific details in the evidence.
Also called FP8 mixed precision.
- sources
- 4
- models
- 2
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 4
Documented in
Further reading
Picked by hand, not extracted: where to read more, not evidence for anything on this page.
- FP8 Formats for Deep Learning (Micikevicius et al., 2022) paper arxiv.org
Evidence
4 spans quoted from the sources, strongest treatment first.
Trained on 27T tokens using FP8 mixed precision
usedoptimizationin MiMo-V2.5-ProXiaomi
Trained on a total of ~48T tokens using FP8 mixed precision.
usedoptimizationin MiMo-V2.5Xiaomi
Trained on a total of ~48T tokens using FP8 mixed precision.
usedtraining objectivein MiMo-V2.5Xiaomi
we adopt an FP8 mixed-precision framework similar to DeepSeek-V3
usedoptimizationin MiMo-V2-FlashXiaomi
Filed alongside
Other methods under optimization :: training precision.
NVFP4BF16NVFP4 pre-trainingFP4+FP8 mixed precisionFP8-precision reinforcement learningMXFP8BF16 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