we experimented with max-based, MSE-based, and Four-Over-Six scaling
1 more from this model
max-based scaling uses the block absolute maximum
specific method · filed under inference & serving
Max-based weight scaling is a quantization scale-selection rule that sets each per-block quantization scale from the block's absolute maximum.1 It is one of several FP4 post-training weight scale-selection strategies, alongside MSE-based and Four-Over-Six scaling.2
Trade-off MSE-based scaling slightly outperforms max in conservative quantization settings, though they are comparable in aggressive settings.3
Summarised by glm-5.3-flash from 2 sources; every sentence is checked against the quote it cites.
max-based scaling uses the block absolute maximum— Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
For weights, we experimented with max-based, MSE-based, and Four-Over-Six— NVIDIA Nemotron 3 Ultra Technical Report
in more conservative quantization settings of 5.03 or 5.43 BPE, MSE slightly outperforms max, whereas in the aggressive quantization setting of 4.85 BPE, they are comparable— NVIDIA Nemotron 3 Ultra Technical Report
Also called Max-based quantization scale calibration, max-based scaling.
2 quotes from 2 sources, one card per model, each opening with its strongest.
we experimented with max-based, MSE-based, and Four-Over-Six scaling
max-based scaling uses the block absolute maximum