# Raw KV Rows vs. Compressed KV Rows

This distinction typically comes up in the context of **KV caching in transformer inference** (and similar key-value storage systems).

## Raw KV Rows

Raw KV rows store the key and value tensors exactly as computed by the attention layers, in full precision:

- **Format**: Full-precision (e.g., FP16, BF16, FP32) tensors
- **Size**: Large — memory grows linearly with sequence length × batch size × layers × heads × head dimension
- **Fidelity**: Exact values as produced during the forward pass
- **Use case**: Default