huyl3-cpu-ComfyUI-PyTorch210Patcher/README.md

ComfyUI PyTorch 2.10 Compatibility Patcher

Auto-fix WanVideoModel for PyTorch 2.10.0+ compatibility

🎯 What This Does

PyTorch 2.10.0 introduced stricter attribute checking that breaks WanVideoModel. This node automatically patches the model class to work with both PyTorch 2.9.0 and 2.10.0+.

🚀 Quick Start

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/huyl3-cpu/ComfyUI-PyTorch210Patcher.git

Usage in ComfyUI

1. Add "PyTorch 2.10 Compatibility Patcher" node to your workflow
2. Connect it BEFORE WanVideo Model Loader
3. Done! ✅

Example Workflow

PyTorch Version Checker (optional)
    ↓
PyTorch 2.10 Compatibility Patcher
    ↓
WanVideo Model Loader
    ↓
... (rest of workflow)

📋 Nodes

1. PyTorch 2.10 Compatibility Patcher

Inputs:

• force_patch (optional): Force patch even if PyTorch < 2.10
• verbose (optional): Print patch status

Outputs:

• status: Patch result message

What it does:

• Detects PyTorch version
• Monkey-patches WanVideoModel.__init__()
• Adds nn.Identity() placeholder for diffusion_model
• Compatible with PyTorch 2.9.0 and 2.10.0+

2. PyTorch Version Checker

Inputs: None

Outputs:

• pytorch_version: e.g. "2.10.0"
• cuda_version: e.g. "12.6"
• needs_patch: True/False

What it does:

• Shows current PyTorch/CUDA versions
• Indicates if patch is needed

🔧 Technical Details

The Problem

# PyTorch 2.10.0 behavior:
class WanVideoModel(BaseModel):
    def __init__(...):
        super().__init__(...)  # ← Calls archive_model_dtypes(self.diffusion_model)
                               # ← AttributeError if diffusion_model doesn't exist!

The Fix

# Our patch:
def patched_init(self, *args, **kwargs):
    self.diffusion_model = nn.Identity()  # ← Placeholder satisfies PyTorch 2.10
    original_init(self, *args, **kwargs)  # ← Now works!
    # Later replaced with actual model in loadmodel()

✅ Compatibility

PyTorch Version	Without Patch	With Patch
2.9.0	✅ Works	✅ Works
2.10.0+	❌ Breaks	✅ Works

📊 Performance

• Zero overhead - patch happens once at workflow start
• No speed impact - same performance as native
• 8-13% faster than PyTorch 2.9.0 (from PyTorch 2.10 improvements)

🆚 Alternative Solutions

Approach	Pros	Cons
This Node	✅ No code changes ✅ Reusable ✅ Easy to use	Requires extra node
Modify WanVideoWrapper	✅ Permanent fix	❌ Modifies original code
Downgrade PyTorch	✅ Simple	❌ Lose 8-13% performance

📝 License

MIT

---

Made for Google Colab A100 80GB optimization 🚀