Skip to content

Conversation

@jhou-jpg
Copy link
Contributor

Summary:
This diff optimizes the entropy warmup logic and should improve its performance with smaller/faster kernels

Since entropy tracks the frequency of appearence for each unique time element, the rounding precision applied on individual latency element can impact the characteristics of entropy convergence. This diff introduces logic to dynamically increase rounding precision to maintain a balance between entropy sensitivity and trend detection

Differential Revision: D87379814

@meta-codesync
Copy link

meta-codesync bot commented Nov 18, 2025

@jhou-jpg has exported this pull request. If you are a Meta employee, you can view the originating Diff in D87379814.

Summary:

This diff fixes an implementation error that was impacting slope and r^2 computations. It also includes optimization to the entropy warmup logic that should improve its performance with smaller/faster kernels. Further numerical stability optimization were brought in from the cross-pr on nvbench NVIDIA/nvbench#286


Since entropy tracks the frequency of appearence for each unique time element, the rounding precision applied on individual latency element can impact the characteristics of entropy convergence. This diff introduces logic to dynamically increase rounding precision to maintain a balance between entropy sensitivity and trend detection

Reviewed By: xuzhao9

Differential Revision: D87379814
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants