Loading your workspace…
Loading your workspace…
One concept at a time
Start with a concept or a question below. Work through the intuition, math, and code at your pace. Ask a follow-up whenever something is unclear.
Already comfortable with transformers?
Due today
Derive how online softmax keeps tiled attention exact. State the running row state, show how a later larger maximum rescales prior contributions, and separate the arithmetic complexity from the HBM-traffic improvement.
Demonstrated, not opened
Concept path
Attention and position must both produce evidence before KV-cache probes enter the queue.
Weakness map
Need a first explanation instead?
Browse the topic library or work through a concept step by step.
Enter to send · Shift+Enter for a new line