DEPLOYMENT_NOTES.md

HuggingFace Space Deployment Notes

Issue: Auto-Rebuild Not Triggering

Problem: Pushing commits to GitHub doesn't always trigger HF Space rebuild automatically.

CRITICAL: HF Space uses main branch, local repo uses master branch!

Correct Push Command:

git push origin master           # Push to GitHub (master branch)
git push hf-new master:main      # Push to HF Space (main branch)

Symptoms:

• Code pushed to GitHub successfully
• Space shows "RUNNING" status
• But still executing old code (verified by same errors)

Solution: Manual Factory Reboot

1. Go to Space Settings: https://huggingface.co/spaces/evgueni-p/fbmc-chronos2/settings
2. Scroll to bottom
3. Click "Factory Reboot"
4. Wait 3-5 minutes

Verifying Rebuild

Check if new code is deployed by looking for these changes in error logs/behavior:

• Old code (pre-7a9aff9): batch_size=256 (default), 9 quantiles
• New code (7a9aff9+): batch_size=32, 3 quantiles

Error signature of old code:

Tried to allocate 10.75 GiB
18.04 GiB memory in use

Expected behavior with new code:

• Much lower memory usage (~3-5 GB instead of 18 GB)
• Should complete 38-border forecast without OOM

Alternative: Direct Space Editing

If repeated rebuilds fail, can edit directly in Space:

1. Go to Space "Files" tab
2. Edit src/forecasting/chronos_inference.py directly
3. Add batch_size=32, quantile_levels=[0.1, 0.5, 0.9] to predict_df() call
4. Space auto-rebuilds on file save

Last Resort: Local Execution

If HF Space continues failing, can run locally:

# Requires: GPU with >8GB VRAM, dataset downloaded
cd /c/Users/evgue/projects/fbmc_chronos2
.venv/Scripts/python.exe evaluate_forecasts.py  # Uses local model

Git Commits Related to This Issue

• 7a9aff9: Memory optimization (batch_size=32, 3 quantiles)
• caf0333: Documentation of Session 11

---

Created: 2025-11-17 Last Updated: 2025-11-17