implementation detail · filed under inference & serving
Selective retention of BF16 precision
Retains BF16 for layers where quantization quality loss outweighed the expected inference-cost benefit.
- source
- 1
- model
- 1
- 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.
BF16 precision for attention linear layers, MoE latent projection layers (i.e, layers for which accuracy degradation outweighed the expected inference-cost benefit after quantization).
corepost trainingin Nemotron 3 UltraNVIDIA
Filed alongside
Other methods under inference & serving :: inference quantization.
FP8 KV-cache quantizationFP8FP4 KV-cache quantizationNVFP4 quantizationFour-Over-SixMXFP4 weight quantizationPost-training quantizationQuantizationBlock-scaled INT8 quantization with stochastic roundingFP8 E4M3 quantizationGGUFMax-based scalingMSE-based scalingMXFP8 activation quantizationNVFP4 KV-cache quantizationNVFP4 ModelOpt re-quantizationSSM cache quantizationAWQ INT4 weight quantization (W4A16)BF16 inferenceBlock-wise E4M3 FP8 weight quantizationChannel-wise quantizationDynamic activation scalingEmbedding and KV-cache quantizationEmpirical bits-per-element budget selection