# Cross modal KV allocation audit

## Description
Audits visual token versus text token KV cache allocation ratios in long multimodal conversations.

## Category
`serving`

## Tags
`serving`, `multimodal`, `kv-cache`, `vram`, `allocation`

## When to Use
Audit multi-turn vision-language models to prevent visual tokens from exhausting available VRAM.

## Operational Limitations
Verifies total VRAM footprint across modality caches; does not compute attention weights.

## 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.