[GH-ISSUE #100] MLA 的实现没有带来任何收益 #91

Closed
opened 2026-10-26 21:07:47 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @XiaoduoAILab on GitHub (Jan 1, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V2/issues/100

下面是 DeepSeekV3 HF官网的MLA实现,可见存入KVCache的数据量,比基线(Llama)还大
befd0b9cfd014cb5f829002f1d1af1b

下面是推理测速,显示速度显著变慢,显存显著增大:
image

Originally created by @XiaoduoAILab on GitHub (Jan 1, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V2/issues/100 下面是 DeepSeekV3 HF官网的MLA实现,可见存入KVCache的数据量,比基线(Llama)还大 ![befd0b9cfd014cb5f829002f1d1af1b](https://github.com/user-attachments/assets/5d169a6d-34e4-45b9-afa1-5c74decdf56c) 下面是推理测速,显示速度显著变慢,显存显著增大: ![image](https://github.com/user-attachments/assets/914eacac-b7c1-44f2-9ed2-b1f482598492)
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-V2#91
No description provided.