Hugging Face Out of Space Fix: The Storage Trap

Short summary

Hugging Face models dump 140GB+ of tensor data into ~/.cache/huggingface by default, which can exhaust small boot drives. The fix is to set HF_HOME environment variable to a secondary storage path before importing any ML libraries, and use huggingface-cli for safe cache management. For read-only production deployments, set HF_HUB_OFFLINE=1 to prevent lock-file write crashes.

• •Set HF_HOME before importing transformers to redirect cache to larger storage
• •Use huggingface-cli scan-cache and delete-cache instead of rm -rf
• •Set HF_HUB_OFFLINE=1 for read-only production inference containers

Generated with AI, which can make mistakes.