제한된 자원으로 언어 모델을 사전학습할 때 fine-grained MoE, Multi-head Latent Attention(MLA), QK-Norm, shared expert와 같은 설계 선택을 고려할 수 있지만, 이들의 효과는 대개 서로 다른 모델 규모·데이터·구현 조건에서 보고되어 직접 일반화하기 어렵다. 본 연구는 A100 4장에서 활성 파라미터(active parameters)를 1.4B로 일치시킨 Dense-GQA, MoE-GQA(전체 8.3B), MoE-MLA(전체 8.4B)를 비교하였다. 세 변형을 한영 이중언어 코퍼스에서 동일한 데이터 순서로 4.59B 토큰까지 학습한 결과, 두 MoE 변형은 Dense-GQA보다 9–10% 낮은 held-out PPL을 보였으며, MoE-GQA와 MoE-MLA의 차이는 1% 미만이었다. 500M-토큰 ablation에서는 QK-Norm 제거, shared expert를 동일 활성 폭의 routed expert로 대체하는 구성, 그리고 RoPE 키의 헤드 간 공유를 검토하였으며, 이 중 QK-Norm 제거가 가장 큰 성능 저하를 보였다. Dense-GQA와 MoE-GQA의 상대적 순위는 seed 반복과 축소된 토큰 예산에서도 유지되었으나, 효과 크기는 예산에 따라 일정하지 않았다. 또한 본 구현에서 MoE 계열의 실측 처리량은 Dense-GQA의 약 23%에 그쳐, 고정 GPU-시간 기준에서는 다른 결론이 가능하다.
Keyword
언어모델, 사전학습, MoE, MLA
KSP Keywords
Fine grained(FG), Multi-head
Copyright Policy
ETRI KSP Copyright Policy
The materials provided on this website are subject to copyrights owned by ETRI and protected by the Copyright Act. Any reproduction, modification, or distribution, in whole or in part, requires the prior explicit approval of ETRI. However, under Article 24.2 of the Copyright Act, the materials may be freely used provided the user complies with the following terms:
The materials to be used must have attached a Korea Open Government License (KOGL) Type 4 symbol, which is similar to CC-BY-NC-ND (Creative Commons Attribution Non-Commercial No Derivatives License). Users are free to use the materials only for non-commercial purposes, provided that original works are properly cited and that no alterations, modifications, or changes to such works is made. This website may contain materials for which ETRI does not hold full copyright or for which ETRI shares copyright in conjunction with other third parties. Without explicit permission, any use of such materials without KOGL indication is strictly prohibited and will constitute an infringement of the copyright of ETRI or of the relevant copyright holders.
J. Kim et. al, "Trends in Lightweight Kernel for Many core Based High-Performance Computing", Electronics and Telecommunications Trends. Vol. 32, No. 4, 2017, KOGL Type 4: Source Indication + Commercial Use Prohibition + Change Prohibition
J. Sim et.al, “the Fourth Industrial Revolution and ICT – IDX Strategy for leading the Fourth Industrial Revolution”, ETRI Insight, 2017, KOGL Type 4: Source Indication + Commercial Use Prohibition + Change Prohibition
If you have any questions or concerns about these terms of use, or if you would like to request permission to use any material on this website, please feel free to contact us
KOGL Type 4:(Source Indication + Commercial Use Prohibition+Change Prohibition)
Contact ETRI, Research Information Service Section
Privacy Policy
ETRI KSP Privacy Policy
ETRI does not collect personal information from external users who access our Knowledge Sharing Platform (KSP). Unathorized automated collection of researcher information from our platform without ETRI's consent is strictly prohibited.
[Researcher Information Disclosure] ETRI publicly shares specific researcher information related to research outcomes, including the researcher's name, department, work email, and work phone number.
※ ETRI does not share employee photographs with external users without the explicit consent of the researcher. If a researcher provides consent, their photograph may be displayed on the KSP.