mirage-project/mirage

[Feature Request] - Performance Improvement for Paged Attention

Open

#428 opened on Jul 27, 2025

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Cuda (232 forks)auto 404
good first issuepersistent kernel

Repository metrics

Stars
 (2,381 stars)
PR merge metrics
 (PR metrics pending)

Description

Feature Description

We can further improve the performance of paged attention tasks with the following optimizations:

  • MAX_TOKENS: MAX_TOKENS is used for preallocating smem; however, current implementation sequentially initialize all tokens' metadata. This should be replaced with dynamic initialization

Contributor guide

Research direction
Understand the current sequential initialization of MAX TOKENS metadata in paged attention, then design a dynamic initialization approach to reduce overhead. Start by examining the smem preallocation and token metadata structure.
Tech stack
ccpp
Domain
backendmachine learning
Issue type
Feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Mostly clear
Prerequisites
CudaGPU programming
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30