KV Cache Explained
Visualizing how transformers avoid recomputing past key/value pairs.
Enter a short sequence (words separated by spaces).
1. Input Tokens
Tokens will appear here.
2. K/V Vector Generation (Current Token)
Waiting for input...
3. Attention Calculation (Using Cache)
The current query attends to its own K/V pairs and all pairs from the cache.
4. KV Cache (Stored Past K/V Pairs)
Stores K/V pairs from past tokens to be reused.
Cache is empty.
Explanation
Click "Start / Next Step" to begin processing the sequence.