Changing the Hugging Face Model Cache Directory (HF_HOME) to Prevent Disk Exhaustion (Error 904)

By Zennith Sys Diagnostics Published 2026-05-22
AIHuggingFacePythonLocalAIStorageDevOps

Are large weights eating up all your disk space? Learn how to override the Hugging Face cache folder configuration and store weights on high-capacity secondary drives.

Dealing with Disk Space Depletion from Large AI Models

Downloading open-weights models (such as stable diffusion, whisper, or llama-3) through Hugging Face compilers quickly fills up system drives. By default, Hugging Face saves downloaded repositories in a hidden user subdirectory (~/.cache/huggingface). On Windows, this is located on your OS drive which can rapidly trigger disk space alerts.

# Step 1: Override Directory Using HF_HOME Variable

You can re-route the downloads globally to a secondary partition or external storage drive by setting environment variables before launching your scripts.

For Linux/macOS: Edit ~/.bashrc or ~/.zscrc and append: `` export HF_HOME="/mnt/storage_drive/huggingface" ` Run source ~/.bashrc to finalize the setting.

For Windows: 1. Open system settings and find "Edit system environment variables". 2. Under User Variables, clicked "New". 3. Set Variable Name to HF_HOME and Variable Value to D:\AI_Models\huggingface (or path to your larger storage drive).

# Step 2: Backwards Compatibility for Old Packages

Older packages might still look for the legacy variable. To be safe, set the older cache variable too:

  • Key: TRANSFORMERS_CACHE
  • Value: Matching paths as defined in HF_HOME`.
  • # Step 3: Clean up and Transfer Existing Weights

    Once variables are confirmed, you can securely cut the previous folder content and paste it into the new directory to avoid re-downloading existing model weights. Delete the previous cache from your boot partition to immediately recover dozens of gigabytes.