Mapping the Schedule x Bit-Width Boundary in Sub-100M Quantisation-Aware Training

The paper's real contribution isn't just that a single schedule works across bit-widths, but that it works across sub-100M models specifically. The implication: quantization-aware training may not require the expensive per-precision tuning that practitioners currently assume, potentially collapsing a source of deployment friction.

This connects directly to the deployment-complete benchmarking work from last week, which exposed the gap between what we measure in labs and what actually works in production. That paper showed benchmark scores often fail to transfer to real deployment contexts. This quantization study matters because it suggests one major source of that transfer failure (precision-specific schedule tuning) may be unnecessary, simplifying the path from research setup to edge deployment. It also echoes the causal methods paper's argument that practitioners waste effort on brute-force hyperparameter search when principled approaches reveal universal patterns.

If teams deploying INT4 models to edge devices report that the 33% warmdown schedule holds without retuning across different model architectures and datasets over the next 6 months, the finding generalizes. If they still need per-precision tweaking in practice, the 1,345-experiment result may reflect lab conditions that don't capture production variability.