v0.1.0: accelerated version with 10X+ speed-up

[VsonicV]

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

---

This discussion was created from the release v0.1.0: accelerated version with 10X+ speed-up.

[Bhoy1]

What is the runtime now with the setup you used earlier? I remember you said 4xH100s and that took about 42 hours. Is it now down to 4 hrs?

[dibbla]

Hi @Bhoy1

The total runtime depends on the number of iterations, but the per-iteration speedup is significant.

On my 4x H100 96GB setup, I'm seeing the iteration time drop from ~770s down to ~63s. That's slightly over a 10x speedup. Of course, your speed up may vary depending on the hardware, but the overall gains are significant.

[Bhoy1]

That's great news! Thank you for that insight