deepseek-ai / deepseek-ai/DeepSpec

[Question] Why DFlash is worse than EAGLE3 for Gemma4-12B in paper's experiment

Open
#76 0 comments 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
7.1k
Forks
667
PR merge metrics
No merged PRs in 30d

Description

Image

In the paper, I notice that in all Qwen series, DFlash is better than EAGLE series, but in Gemma4 the DFlash's acceptance is always worse than EAGLE3. The paper does not explain why, just saying that DSpark is getting consistant gain over different models. I'm quite curious why the DFlash failed to beat EAGLE3 in Gemma4? Is it pure model related issues? Looking forward to getting some applies🤔

Contributor guide

No contributing guide indexed for this repository

Research direction

The issue concerns the paper’s Gemma4-12B experiment and compares DFlash, EAGLE3, DSpark, and Qwen results. Start by reading the paper’s experiment description and examining the reported acceptance results; done means a documented, evidence-based explanation of the model-specific difference.

Written by the indexing model from the issue text.

Assessment

Domain
machine-learning
Issue type
Documentation
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.