deepseek-ai / deepseek-ai/FlashMLA

Can MHA be used in the DSA prefill stage?

Open
#125 1 comment 1 reaction 0 assignees View on GitHub
Dominant language
C++
Stars
12.9k
Forks
1.2k
Avg merge
4h 20m
Merged PRs (30d)
2

Description

The current sparse flash MLA implements the MQA version, where the last dimension of q and k increases from 192 to 576, adding computational overhead. Is it possible to replace MQA with MHA to reduce the computational load of sparse flash MLA during the prefill stage?

Contributor guide

No contributing guide indexed for this repository

Research direction

Start by tracing the sparse flash MLA implementation for the DSA prefill stage and compare the current MQA formulation with the proposed MHA approach. Determine whether MHA can reduce computational load without changing the required attention results; done means the feasibility and measured trade-offs are documented or implemented with validation.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
machine-learning, performance
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.