This poster includes a walkthrough of DeepSeek’s multihead latent attention with captions, and a detailed comparison between various forms of attention, including the required sizes of kv caches, and the 3d model of each type of attention head.

The matrix images are actually from the real deepseek model - mostly the weights from the first layer of DeepSeek V3.

The poster looks great in an simple frame you can pick up on amazon, and is a great way to see how MLA works and just a nice way to decorate your walls.

• Beautifully printed high resolution poster

• Printed on high quality Canon Pro Luster paper for excellent color reproduction and longevity

• Actual printed size is slightly smaller (~12 7/8” x 18 7/8”), fits well in most standard 13x19” frames, example here.

• Frame not included

• Only shipping to US addresses currently, please allow 3-8 business days for shipping & handling