Attention Internals

The full softmax(QK^T / sqrt(d)) V derivation, why the sqrt(d) scaling, RoPE and relative position, causal masking.

0 / 10 solved
0/10 solved
Question Tags Difficulty Success
1. Decode-Step Attention Shapes
attentiontransformers +3
Medium —
2. No Peeking Ahead
attentiontransformers +1
Medium —
3. The Saturation Trap
attentiontransformers +1
Medium —
4. Fewer Keys, Same Heads
gqakv-cache +1
Medium —
5. Spin to Encode
attentiontransformers +1
Hard —
6. Position Encoding Showdown
alibirope +1
Hard —
7. KV Bytes Per Token: MHA, GQA, MLA
mlakv-cache +3
Hard —
8. The Sink That Never Drains
attention-sinksstreamingllm +1
Hard —
9. Split the KV, Not the Query
flash-decodingmemory-bound +1
Hard —
10. NTK-by-Parts
yarnrope +1
Insane —

Up next

10 of 10 questions here are still unsolved.