Pseudo-Lab / EfficientLLM

Repository from Github https://github.comPseudo-Lab/EfficientLLMRepository from Github https://github.comPseudo-Lab/EfficientLLM

EfficientLLM: Speed always wins

PseudoLab Discord Community Stars Badge Forks Badge Pull Requests Badge Issues Badge GitHub contributors

speed always wins

๐Ÿš€ EfficientLLM: Speed always wins repository์— ์˜ค์‹  ๊ฒƒ์„ ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค! ์ €ํฌ๋Š” Transformer ์•„ํ‚คํ…์ฒ˜์˜ ๊ทผ๋ณธ์ ์ธ ๋น„ํšจ์œจ์„ฑ์„ ํƒ๊ตฌํ•˜๊ณ , Sparse Attention๊ณผ Speculative Decoding ๊ฐ™์€ ์ตœ์‹  ์ตœ์ ํ™” ๊ธฐ์ˆ ๋“ค์„ ๊นŠ์ด ์žˆ๊ฒŒ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” Large Language Models์˜ ์„ฑ๋Šฅ ์žฅ๋ฒฝ์„ ๋ŒํŒŒํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. LLM์„ ๋” ๋น ๋ฅด๊ณ , ๋” ํšจ์œจ์ ์ด๋ฉฐ, ๋” ์‰ฝ๊ฒŒ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก ๋งŒ๋“œ๋Š” ์—ฌ์ •์— ํ•จ๊ป˜ํ•ด์ฃผ์„ธ์š”!

๐ŸŒŸ ํ”„๋กœ์ ํŠธ ๋ชฉํ‘œ (Project Vision)

"์•„ํ‚คํ…์ฒ˜ ์ˆ˜์ค€์˜ ๊นŠ์ด ์žˆ๋Š” ์ดํ•ด๋ฅผ ํ†ตํ•ด LLM ์ถ”๋ก ์˜ ํ˜„์‹ค์ ์ธ ์žฅ๋ฒฝ์„ ๋„˜์–ด์„œ๋‹ค"

Transformer์˜ $O(N^2)$ ๋ณต์žก๋„, ๋ง‰๋Œ€ํ•œ ๋ฉ”๋ชจ๋ฆฌ ์š”๊ตฌ๋Ÿ‰์€ Long-Context, AI Agent์™€ ๊ฐ™์€ ์ฐจ์„ธ๋Œ€ AI ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๊ฐ€์žฅ ํฐ ๋ณ‘๋ชฉ์ž…๋‹ˆ๋‹ค. ์ˆ˜ ์กฐ์›์˜ ๋ฐ์ดํ„ฐ์„ผํ„ฐ, ์ˆ˜ ์–ต์›์˜ ์„œ๋น™ ๋น„์šฉ์€ ํ˜์‹ ์ ์ธ LLM ๊ฐœ๋ฐœ๊ณผ ์„œ๋น„์Šค๋ฅผ ๊ฐ€๋กœ๋ง‰๋Š” ํ˜„์‹ค์ ์ธ ์žฅ๋ฒฝ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

๋ณธ ํ”„๋กœ์ ํŠธ๋Š” LLM ์ถ”๋ก  ํšจ์œจ์„ ๋†’์ด๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์ถ•, Sparse Attention๊ณผ Speculative Decoding์„ ์ค‘์‹ฌ์œผ๋กœ ์ตœ์‹  ์—ฐ๊ตฌ๋“ค์„ ํƒ๊ตฌํ•˜์—ฌ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์—ญ๋Ÿ‰์„ ๊ฐ–์ถ”๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.

  • ํ•ต์‹ฌ ์›๋ฆฌ ์ดํ•ด: ๊ฐ ์ตœ์ ํ™” ๊ธฐ์ˆ ์˜ ์ž‘๋™ ๋ฐฉ์‹์„ ์•„ํ‚คํ…์ฒ˜ ์ˆ˜์ค€์—์„œ ๊นŠ์ด ์žˆ๊ฒŒ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค.
  • ํ†ต์ฐฐ๋ ฅ ํ™•๋ณด: ์–ด๋–ค ์ƒํ™ฉ์—์„œ ์–ด๋–ค ๊ธฐ์ˆ ์ด ํšจ๊ณผ์ ์ธ์ง€์— ๋Œ€ํ•œ ํ†ต์ฐฐ๋ ฅ์„ ๊ธฐ๋ฆ…๋‹ˆ๋‹ค.
  • ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ: ๋น„์šฉ๊ณผ ์†๋„์˜ ์ œ์•ฝ์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ์‹ค์งˆ์ ์ธ ์—ญ๋Ÿ‰์„ ๊ฐ–์ถฅ๋‹ˆ๋‹ค.
  • ์ง€์‹ ๊ณต์œ : ๋ชจ๋“  ํ•™์Šต ๊ฒฐ๊ณผ๋ฌผ์„ ๊ณต๊ฐœํ•˜์—ฌ ๊ตญ๋‚ด LLM ์ƒํƒœ๊ณ„์— ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค.

๐Ÿง‘ ์—ญ๋™์ ์ธ ํŒ€ ์†Œ๊ฐœ (Dynamic Team)

์—ญํ•  ์ด๋ฆ„ LinkedIn
Project Manager ์ „๊ฒฝํ˜ธ LinkedIn
Member ๊ธธ์žฌ์€ LinkedIn
Member ๊น€์Šน์šฐ LinkedIn
Member ๊น€ํ˜•๊ท  LinkedIn
Member ๋ฐ•์žฌ์šฑ LinkedIn
Member ์ด์Šน์•„ LinkedIn

๐Ÿ’ป ์ฃผ์ฐจ๋ณ„ ํ™œ๋™ (Activity History)

  • ์‹œ๊ฐ„: ๋งค์ฃผ ํ™”์š”์ผ 20:00-22:00
  • ์žฅ์†Œ: Room-AT
๋‚ ์งœ ๋‚ด์šฉ ๋ฐœํ‘œ์ž ์˜์ƒ
2025/9/9 OT ์ „๊ฒฝํ˜ธ
2025/9/16 Speed Always Wins: A Survey on Efficient Architectures for Large Language Models
Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding
์ „๊ฒฝํ˜ธ
๊ธธ์žฌ์€
2025/9/23 vLLM ๋ฐ•์žฌ์šฑ
์ด์Šน์•„
2025/9/30 ๊น€์Šน์šฐ
๊น€ํ˜•๊ท 
2025/10/7 ์ „๊ฒฝํ˜ธ
๊ธธ์žฌ์€
2025/10/14 ๋ฐ•์žฌ์šฑ
์ด์Šน์•„
2025/10/21 ๊น€์Šน์šฐ
๊น€ํ˜•๊ท 
2025/10/30 ์ „๊ฒฝํ˜ธ
๊ธธ์žฌ์€
2025/11/4 ๋ฐ•์žฌ์šฑ
์ด์Šน์•„
2025/11/11 ๊น€์Šน์šฐ
๊น€ํ˜•๊ท 
2025/11/18 ์ „๊ฒฝํ˜ธ
๊ธธ์žฌ์€
2025/11/25 ๋ฐ•์žฌ์šฑ
์ด์Šน์•„
2025/12/2 ๊น€์Šน์šฐ
๊น€ํ˜•๊ท 
2025/12/9 ์ „๊ฒฝํ˜ธ
๊ธธ์žฌ์€
2025/12/16 ๋ฐ•์žฌ์šฑ
์ด์Šน์•„
2025/12/23 ๊น€์Šน์šฐ
๊น€ํ˜•๊ท 

๐Ÿ’ก ํ•™์Šต ์ž์› (Learning Resources)

ํ•ต์‹ฌ Survey ๋…ผ๋ฌธ

๋…ผ๋ฌธ ํƒ์ƒ‰์„ ์œ„ํ•œ ๋ ˆํฌ์ง€ํ† ๋ฆฌ

๐ŸŒฑ ์ฐธ์—ฌ ์•ˆ๋‚ด (How to Engage)

  • ๋นŒ๋”๋กœ ์ฐธ์—ฌ โ€” ํ”„๋กœ์ ํŠธ ๊ธฐํšยท์šด์˜ ์ฃผ๋„
  • ๋Ÿฌ๋„ˆ๋กœ ์ฐธ์—ฌ โ€” ์—ฐ๊ตฌยท๊ฐœ๋ฐœยทํ…Œ์ŠคํŠธ ๋“ฑ ์‹คํ–‰
  • ์ฒญ๊ฐ• ์ฐธ์—ฌ โ€” ๊ณต๊ฐœ ์„ธ์…˜ ์ฐธ์—ฌ ๊ฐ€๋Šฅ

โ—๏ธ์ฐธ์—ฌ ๋งํฌ: ๊ฐ€์งœ์—ฐ๊ตฌ์†Œ ๋””์Šค์ฝ”๋“œ โ—๏ธ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ์ฑ„๋„: ๋””์Šค์ฝ”๋“œ #{{์ฑ„๋„๋ช…}}

๋ˆ„๊ตฌ๋‚˜ ์ฒญ๊ฐ•์„ ํ†ตํ•ด ๋ชจ์ž„์„ ์ฐธ์—ฌํ•˜์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

  1. ํŠน๋ณ„ํ•œ ์‹ ์ฒญ ์—†์ด ์ •๊ธฐ ๋ชจ์ž„ ์‹œ๊ฐ„์— ๋งž์ถ”์–ด ๋””์Šค์ฝ”๋“œ #Room-GH ์ฑ„๋„๋กœ ์ž…์žฅ
  2. Magical Week ์ค‘ ํ–‰์‚ฌ์— ์ฐธ๊ฐ€
  3. Pseudo Lab ํ–‰์‚ฌ์—์„œ ๋งŒ๋‚˜๊ธฐ

Acknowledgement ๐Ÿ™

์ด ํ”„๋กœ์ ํŠธ๋Š” ๊ฐ€์งœ์—ฐ๊ตฌ์†Œ Open Academy๋กœ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ๋ถ„์˜ ์ฐธ์—ฌ์™€ ๊ธฐ์—ฌ๊ฐ€ โ€˜์šฐ์—ฐํ•œ ํ˜๋ช…(Serendipity Revolution)โ€™์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋‘์—๊ฒŒ ๊นŠ์€ ๊ฐ์‚ฌ๋ฅผ ์ „ํ•ฉ๋‹ˆ๋‹ค. OOO is developed as part of Pseudo-Lab's Open Research Initiative. Special thanks to our contributors and the open source community for their valuable insights and contributions.

About Pseudo Lab ๐Ÿ‘‹๐Ÿผ

Pseudo-Lab is a non-profit organization focused on advancing machine learning and AI technologies. Our core values of Sharing, Motivation, and Collaborative Joy drive us to create impactful open-source projects. With over 5k+ researchers, we are committed to advancing machine learning and AI technologies.

Contributors ๐Ÿ˜ƒ



License ๐Ÿ—ž

This project is licensed under the MIT License.

About