ES-at-Scale Library

This release introduces the new ES-at-Scale library with the following features:

1. A more extensible coding interface for easier adaptation to new tasks/models. @conorfhayes
2. A refined infrastructure with more stable training and improved precision handling. @dibbla @conorfhayes @kschweig @VsonicV
3. Support for the MATH benchmarks. @conorfhayes
4. A detailed user guide/documentation. @conorfhayes @VsonicV

v0.1.0: Accelerated Version with 10X+ Speed-up

This initial release includes three versions of ES implementations:

• Original version: the source codes to reproduce the experimental results of ES in the original paper. @VsonicV @sunrainyg
• Accelerated version: an accelerated version using vLLM to achieve 10X+ speed-up in running time. @dibbla @VsonicV
• i.i.d. version: similar to the original version, but with complete i.i.d. noise for model perturbations. @VsonicV