Expand description
§Mixed-Precision Policy
Assigns per-layer bit-widths to meet a target average bits-per-parameter budget, while respecting per-layer sensitivity. More sensitive layers receive higher bit-widths.
§Greedy algorithm
- Initialise every layer to the minimum bit-width.
- While the current average bits < target: a. Find the layer whose upgrade (to the next bit-width) yields the largest marginal sensitivity reduction per extra bit spent. b. Upgrade that layer.
- Return the final assignment.
Structs§
- Mixed
Precision Policy - Per-layer bit-width assignment produced by the greedy sensitivity policy.