# Multi-Head Latent Attention Audit

## Description
Calculates DeepSeek Multi-Head Latent Attention (MLA) low-rank compression factors and verifies KV cache reduction relative to MHA and GQA.

## Category
`architecture`

## Tags
`mla`, `deepseek`, `attention`, `kv-cache`, `compression`

## When to Use
Audit architectural memory savings of models using low-rank key/value compression (DeepSeek V2/V3/R1).

## Operational Limitations
Pure mathematical verification of KV projection tensor sizes and per-token memory footprint.

## Verification Script
Run the deterministic verification checker:
```bash
python3 scripts/check.py examples/synthetic.json
```
Exit codes:
- `0`: Check passed and thresholds satisfied.
- `1`: Threshold breach or operational violation detected.
- `2`: Schema error or invalid usage.