How to Diagnose ComfyUI Dependency Conflicts

The Problem Nobody Warns You About

You install ComfyUI. It works. You add 10 custom nodes. Still works. You add the 11th — and suddenly nothing works. PyTorch is gone. CUDA is gone. Your GPU is a paperweight.

This isn't a bug. It's dependency conflict — the #1 hidden killer of ComfyUI environments. Each plugin has its own requirements.txt, and when you install 20+ plugins, their demands often contradict each other.

Real Examples We See Every Week

• Plugin A wants numpy>=2.0, Plugin B wants numpy<2.0
• Plugin C wants pillow>=10.3, Plugin D wants pillow<10.0
• BizyAir auto-updates on startup and silently changes your starlette version
• Installing nunchaku without --no-deps pulls in a different PyTorch build

Step 1: Find the Conflict

Check startup logs

Look for IMPORT FAILED in the ComfyUI console:

IMPORT FAILED: ComfyUI-TeaCache
  AttributeError: module 'comfy.model_sampling' has no attribute '_precompute_freqs_cis'

Run pip check

pip check

This lists all broken dependency relationships.

Check Manager's failure endpoint

Visit http://127.0.0.1:8188/v2/customnode/import_fail_info_bulk for a JSON list of failed imports.

Step 2: Classify — Not Every Warning Is Real

Step 3: Fix Without Breaking More

Protect core runtime: torch, torchvision, numpy, pillow must never be accidentally downgraded.

# Bad
pip install -U transformers

# Good
pip install transformers==4.57.6

Use --no-deps for dangerous packages. Use pre-built wheels for nunchaku, insightface, etc.

Common Recipes

• BizyAir: pip install starlette==0.37.2 + BIZYAIR_SKIP_UPDATE=1
• HuggingFace stack: Install diffusers, transformers, huggingface-hub, accelerate together
• numpy war: Keep numpy==1.26.4, accept pip check warnings

When All Else Fails

Reinstall core packages to known-good versions, or start fresh. Or book a remote fix session — we untangle dependency hell in under 30 minutes.

Full reference: Dependency Conflicts docs.