Baseten's post states that NVFP4 represents values in 4 bits and divides tensors into blocks of 16 values, each with its own scale. It also explains that activation scales require calibration data because activations depend on model inputs, while weights are fixed. These details make the quantization recipe a measurement and allocation problem, not merely a uniform format conversion.
Quantization choices should be judged as a whole recipe, not by ranking layers in isolation. This is a vendor-authored technical explanation; it does not establish independent benchmark results or prove that SaturationQuant is best across models and workloads.
Thesis
High-quality NVFP4 quantization depends on choosing precision per layer, and estimating those choices requires accounting for interactions among quantized layers.
Observed facts
The post says NVFP4 divides tensors into blocks of 16 values, each with its own scale.
1The post says activation scales require calibration data because activations depend on inputs, while weights are fixed.
1The post describes SaturationQuant using Aumann–Shapley scoring to account for interactions among quantized layers.
1
Mechanism
The post contrasts architecture-based heuristics with sensitivity scoring. Isolated-layer scores measure the effect of quantizing one layer while other layers remain at baseline; Baseten argues those scores cannot simply be added for a mixed-precision recipe because loss can saturate as more layers are quantized. SaturationQuant uses Aumann–Shapley scoring to build this effect into scores that can be summed.
Why now
NVFP4 offers smaller weights and potentially faster computation on hardware with native FP4 support, so preserving quality while assigning precision has direct deployment relevance.
Implications
Quantization choices should be judged as a whole recipe, not by ranking layers in isolation.
Calibration data and the target workload matter when assigning precision.
Evidence boundary
The source explains the method and its rationale, but the supplied material does not provide independent comparative validation.
Watch conditions
Independent evaluations across models and representative calibration datasets would show whether the described scoring approach predicts quality reliably in practice.