« Attention clairsemée » : différence entre les versions


Aucun résumé des modifications
Aucun résumé des modifications
 
Ligne 21 : Ligne 21 :
'''native sparse attention'''
'''native sparse attention'''
   
   
<!-- Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance.-->
Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance.
 
==Sources==
==Sources==
[https://espace.etsmtl.ca/id/eprint/3299/ Aroosa Hameed (2023) - attention clairsemée ]
[https://espace.etsmtl.ca/id/eprint/3299/ Aroosa Hameed (2023) - attention clairsemée ]
Ligne 29 : Ligne 30 :
[https://aarnphm.xyz/thoughts/papers/DeepSeek_V3_2.pdf  DeepSeek - sparse attention]
[https://aarnphm.xyz/thoughts/papers/DeepSeek_V3_2.pdf  DeepSeek - sparse attention]


[[Catégorie:Publication]]
[[Catégorie:GRAND LEXIQUE FRANÇAIS]]

Dernière version du 19 avril 2026 à 12:32

Définition

Technique d'optimisation d'une architecture auto-attentive qui vise à réduire le nombre de calculs nécessaires au mécanisme d'auto-attention.

Compléments

Au lieu de comparer tous les segments textuels entre eux, un processus de sélection concentre l'attention sur certains segments en particulier selon différentes heuristiques. Par exemple, es heuristiques exploitent une largeur de fenêtre, la position d'un jeton dans une phrase, la similarité des segments, etc.

Cette technique d'optimisation a été mise de l'avant par la société DeepSeek-AI.

Français

attention clairsemée

attention parcimonieuse

attention creuse

attention clairsemée native

Anglais

sparse attention

native sparse attention

Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance.

Sources

Aroosa Hameed (2023) - attention clairsemée

Wikipedia - attention clairsemée

DeepSeek - sparse attention

Contributeurs: Patrick Drouin, wiki