Research 0127: Tiny-AI quantisation doc status¶
- Date: 2026-05-15
- Workstream: tiny-AI documentation hygiene
- Tags: tiny-ai, quantisation, docs, fork-local
Question¶
Does docs/ai/quantization.md still describe the shipped tiny-AI quantisation state accurately after the v3/v4 PTQ follow-ups landed?
Findings¶
The central quantisation page already listed learned_filter_v1 and nr_metric_v1 as dynamic-PTQ models, but its caveat section still said that no model was currently quantised. That sentence is stale: the registry contains non-fp32 rows for learned_filter_v1, nr_metric_v1, vmaf_tiny_v3, and vmaf_tiny_v4, each with quant_mode: "dynamic", an int8_sha256, and a 0.01 PLCC budget.
The per-model cards for vmaf_tiny_v3 and vmaf_tiny_v4 already carry the measured dynamic-PTQ drops and file sizes:
| Model | fp32 | int8 | Drop |
|---|---|---|---|
vmaf_tiny_v3 | 4 496 B | 4 267 B | 0.000120 |
vmaf_tiny_v4 | 14 046 B | 7 769 B | 0.000145 |
No registry row currently uses static PTQ or QAT, so the accurate operator-facing caveat is "all shipped int8 sidecars are dynamic PTQ", not "no model is currently quantised".
Decision Matrix¶
| Option | Pros | Cons | Decision |
|---|---|---|---|
| Leave the page as-is | No churn | Direct contradiction with registry and model cards | Rejected |
| Update only the caveat | Removes the contradiction | Still omits the shipped v3/v4 sidecars from the summary table | Rejected |
| Update the table and caveat | Aligns the central page with the registry and model cards | Documentation-only PR | Accepted |