Skip to content

Net-energy Expert Controller for Test-time Adaptive Routing (Nectar)

Notifications You must be signed in to change notification settings

AbiralShakya/Nectar

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

86 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Nectar

Nectar considers energy of the mixture of experts model (fundamental to scaling for modern LLMs, Diffusion models, etc) and reroutes experts during inference (a new paradigm in test time adaption techniques) based upon energy profiles.

to consider: kernel optimizations for dequantization & ttt specifically (fused kernels, large chunk updates, tile packing), look into TTT architecture and MIT paper: https://arxiv.org/pdf/2505.23884 (Test time training done right) to consider (convexity proofs annd logic, large chun)

Nectar is essentially an analogy to switch transformer but hardware aware (for now), or another interpretation is considering TTT research on model weight updates but now as larger amount of experts are being intergrated in next gen LLM and GenAI, bringing that test time self supervised learning to re routing itself.

TODO: consider model sharding within GPU for memory management (good experimentation area)

THIS IS IMPORTANT: "Hardware-aware memory mapping" from Hyper Accel Adelia (https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11075108) --> understand fully & see how to implement for decode phase for nectar !

About

Net-energy Expert Controller for Test-time Adaptive Routing (Nectar)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published