|
|
--- |
|
|
license: mit |
|
|
language: |
|
|
- en |
|
|
--- |
|
|
# 🧠 WAN2GP Video Generator |
|
|
|
|
|
A reproducible pipeline for generating videos using custom LoRA modules, TorchInductor, and CUDA 12.8. Built for portability, disaster recovery, and clean builds across systems. |
|
|
|
|
|
## 🖼️ Sample Output |
|
|
|
|
|
Here’s a successful run showing model pinning, async shuttle setup, and Torch compilation: |
|
|
|
|
|
 |
|
|
|
|
|
## 🧱 Environment Setup |
|
|
|
|
|
This project requires a clean environment with **Python 3.10**, **CUDA 12.8**, and **Visual Studio Build Tools 2022**. |
|
|
|
|
|
### 1. 🔧 Install Visual Studio Build Tools 2022 |
|
|
|
|
|
Required to compile TorchInductor kernels, `xformers`, and transformer modules. |
|
|
|
|
|
- Download: [Visual Studio Build Tools 2022](https://visualstudio.microsoft.com/visual-cpp-build-tools/) |
|
|
- Enable workload: ✅ **Desktop development with C++** |
|
|
- Launch from: **x64 Native Tools Command Prompt for VS 2022** |
|
|
|
|
|
### 2. 🐍 Install Miniconda |
|
|
|
|
|
- Download: [Miniconda for Windows](https://docs.conda.io/en/latest/miniconda.html) |
|
|
|
|
|
### 3. 🧪 Create the Conda Environment |
|
|
|
|
|
*(Run all commands in the x64 Native Tools Command Prompt for VS 2022)* |
|
|
|
|
|
``` |
|
|
conda create -n wan2gp python=3.10 |
|
|
conda activate wan2gp |
|
|
``` |
|
|
|
|
|
### 4. 📥 Download WAN2GP from GitHub |
|
|
|
|
|
Clone the base pipeline from [deepbeepmeep's GitHub](https://github.com/deepbeepmeep/wan2gp): |
|
|
|
|
|
``` |
|
|
git clone https://github.com/deepbeepmeep/wan2gp.git |
|
|
cd wan2gp |
|
|
pip install -r requirements.txt |
|
|
``` |
|
|
|
|
|
## 📦 Step-by-Step Installation from Custom Wheels |
|
|
|
|
|
*(Continue in the same x64 Native Tools Command Prompt for VS 2022)* |
|
|
|
|
|
Install the following wheels **in this exact order** to ensure compatibility. |
|
|
|
|
|
#### 1. Download Wheels via Browser |
|
|
|
|
|
Visit the wheels_for_windows directory and **manually download each** `.whl` **file** to a local folder (e.g., `C:\wan2gp\wheels_for_windows`). (right click and save as) |
|
|
|
|
|
#### 2. Install Wheels Locally |
|
|
|
|
|
Open your **x64 Native Tools Command Prompt for VS 2022**, activate your conda enviornment, navigate to the folder containing the wheels, and run: |
|
|
|
|
|
``` |
|
|
pip install triton_windows-3.4.0.post20-cp310-cp310-win_amd64.whl |
|
|
pip install torch-2.9.0.dev20250909-cu128-cp310-cp310-win_amd64.whl |
|
|
pip install torchvision-0.24.0.dev20250909-cu128-cp310-cp310-win_amd64.whl |
|
|
pip install torchaudio-2.8.0.dev20250909-cu128-cp310-cp310-win_amd64.whl |
|
|
pip install flash_attn-2.8.3-cp310-cp310-win_amd64.whl |
|
|
pip install xformers-0.0.33-f2043594.d20251008-cp39-abi3-win_amd64.whl |
|
|
``` |
|
|
|
|
|
|
|
|
## 🔧 Update [`wgp.py`](https://wgp.py) and Enable Compilation |
|
|
|
|
|
Before launching, follow these steps to enable Torch compilation and apply your patched [`wgp.py`](https://wgp.py): |
|
|
|
|
|
### For WAN2GP Version 8.995: |
|
|
|
|
|
1. Launch the program once: |
|
|
|
|
|
``` |
|
|
python wgp.py |
|
|
``` |
|
|
|
|
|
2. Go to **Configuration → Performance** |
|
|
|
|
|
3. Enable ✅ **Compile Transformers** |
|
|
|
|
|
4. Click **Apply Settings**, then exit the program |
|
|
|
|
|
5. Overwrite the default [`wgp.py`](https://wgp.py) with your patched version (from this repo) |
|
|
|
|
|
6. Relaunch: |
|
|
|
|
|
``` |
|
|
python wgp.py |
|
|
``` |
|
|
|
|
|
### For WAN2GP Versions **Newer Than 8.995**: |
|
|
|
|
|
Add the following lines to the **top of [`wgp.py`](https://wgp.py)**: |
|
|
|
|
|
``` |
|
|
import torch._dynamo |
|
|
torch._dynamo.config.accumulated_recompile_limit = 512 # or higher |
|
|
torch._dynamo.config.verbose = True |
|
|
torch._dynamo.config.suppress_errors = True |
|
|
|
|
|
def compile_or_fallback(model, example_inputs): |
|
|
try: |
|
|
print("🧪 Attempting Torch compile...") |
|
|
compiled_model = torch.compile(model, backend="inductor") |
|
|
try: |
|
|
compiled_model(*example_inputs) |
|
|
except Exception as runtime_error: |
|
|
print("⚠️ Runtime error during dry run. Falling back to eager mode.") |
|
|
print("Runtime error:", runtime_error) |
|
|
return model |
|
|
print("✅ Compilation succeeded.") |
|
|
return compiled_model |
|
|
except Exception as compile_error: |
|
|
print("⚠️ Compilation failed. Falling back to eager mode.") |
|
|
print("Compile error:", compile_error) |
|
|
return model |
|
|
``` |
|
|
|
|
|
This ensures graceful fallback and verbose diagnostics during Torch compilation. |
|
|
|
|
|
## 🔁 Upgrading WAN2GP Code |
|
|
|
|
|
If you upgrade the WAN2GP codebase (e.g., by pulling updates from GitHub): |
|
|
|
|
|
- ✅ Re-run `pip install -r `[`requirements.txt`](https://requirements.txt) to install any new dependencies |
|
|
- ⚠️ Then **reinstall all custom wheels** in the same order as above — upgrading may overwrite them |
|
|
- 🧩 Reapply your patched [`wgp.py`](https://wgp.py) unless deepbeepmeep integrates these changes upstream |
|
|
|
|
|
## 🚀 Launch Instructions |
|
|
|
|
|
Launch the pipeline using: |
|
|
|
|
|
``` |
|
|
python wgp.py |
|
|
``` |
|
|
|
|
|
**Important**: Always run from the **x64 Native Tools Command Prompt for VS 2022** to ensure compiler visibility ([`cl.exe`](https://cl.exe)) and proper environment variables. |
|
|
|
|
|
## 🧠 Runtime Behavior |
|
|
|
|
|
- TorchInductor compiles kernels on first run - this takes a few minutes, be patient |
|
|
- LoRA modules are injected and pinned to RAM |
|
|
- Async shuttles handle memory transfers across GPU and CPU |
|
|
- Output videos saved to `outputs/` with timestamped filenames |
|
|
|
|
|
## ⚠️ Optional Windows Tweaks (for long async runs) |
|
|
|
|
|
If you encounter `WinError 10055` (socket buffer exhaustion), apply these registry tweaks: |
|
|
|
|
|
| Registry Key | Type | Value | |
|
|
| ------------------- | ----- | ------- | |
|
|
| `MaxUserPort` | DWORD | `65534` | |
|
|
| `TcpTimedWaitDelay` | DWORD | `30` | |
|
|
|
|
|
Location: |
|
|
`HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\Tcpip\Parameters` |
|
|
|
|
|
Reboot required after applying. |
