Real-world uses

Created for this library

1. 1.

   A research team experiments with model width up to 4096 in its dense hidden layers to find the right capacity for its dataset.

2. 2.

   An ML platform team caps model width on its on-device model to keep latency and memory within budget.

3. 3.

   A startup picks model width to fit the inference budget rather than maximum accuracy because per-call cost dictates the production design.