Cache Sessions Guide

This guide explains cache utilities available in AMMM for PyTensor compilation artifacts.

What This Guide Covers

• ammm-cache CLI usage
• SessionManager for explicit cache isolation in custom scripts
• Practical cache hygiene

1. ammm-cache CLI

AMMM provides a cache management CLI:

ammm-cache info
ammm-cache prune
ammm-cache clear

Default cache root:

• ~/.ammm/compiled

Useful options:

ammm-cache info --verbose
ammm-cache prune --max-size 3000 --max-count 20 --max-age 14
ammm-cache prune --dry-run --verbose
ammm-cache clear --force
ammm-cache info --cache-root /custom/path

2. Session Isolation in Python Scripts

For advanced workflows, you can isolate compilation caches per config hash using SessionManager.

from pathlib import Path
from src.utils.session_manager import SessionManager

config = {
    "media": [{"spend_col": "tv_spend"}, {"spend_col": "search_spend"}],
    "extra_features_cols": ["competitor_sales"],
    "ad_stock_max_lag": 8,
}

with SessionManager(config=config, cache_root=Path.home() / ".ammm" / "compiled", verbose=True):
    # Run model fitting code here
    pass

This is most useful when you run many shape-varying configurations in one Python process.

3. Important Clarification

Current runme.py does not expose a documented AMMM_CACHE_MODE flag. Treat cache session management as:

• automatic runtime behaviour handled internally where applicable; and/or
• explicit control in custom scripts via SessionManager.

4. Recommended Hygiene

1. Use ammm-cache info weekly if you run many experiments.
2. Use ammm-cache prune with bounded size/age.
3. Use clear --force before benchmark reruns if you need a cold start.

5. Troubleshooting

• If cache directories grow unexpectedly, verify --cache-root.
• If custom script runs exhibit shape reuse issues, use explicit SessionManager contexts per configuration family.