ZEROGPU_SETUP_GUIDE.md

Additional Environment Configuration for ZeroGPU

Add this to your Hugging Face Space's Settings → Variables:

Environment Variables

Required:

ZEROGPU_OFFLOAD_DIR=/tmp/zerogpu-offload

Recommended:

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
CUDA_LAUNCH_BLOCKING=0
HF_HUB_ENABLE_HF_TRANSFER=1

Alternative: Direct Folder Creation

If the above doesn't work, you can also try creating a startup script in your Space.

Space Configuration File

Create or modify your Space's README.md to include:

---
title: Wan2.2-Fast-I2I
emoji: 💻
colorFrom: purple
colorTo: gray
sdk: gradio
sdk_version: 5.44.1
app_file: app.py
pinned: false
hardware: a10g-large

The hardware: a10g-large ensures you get a ZeroGPU instance with sufficient memory.

Dockerfile Alternative

If you need more control, create a Dockerfile:

FROM python:3.10

# Create offload directory
RUN mkdir -p /data-nvme/zerogpu-offload && chmod 755 /data-nvme/zerogpu-offload

# Set environment variables
ENV ZEROGPU_OFFLOAD_DIR=/data-nvme/zerogpu-offload
ENV PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

# Install your requirements
COPY requirements.txt .
RUN pip install -r requirements.txt

# Copy your app
COPY . /app
WORKDIR /app

CMD ["python", "app.py"]

Testing the Fix

The modifications I made to app.py should handle:

1. ✅ Automatic directory creation - Creates /data-nvme/zerogpu-offload or falls back to /tmp/zerogpu-offload
2. ✅ Permission handling - Gracefully handles cases where NVMe isn't writable
3. ✅ Environment variables - Sets proper PyTorch memory configuration
4. ✅ ZeroGPU decorators restored - Keeps @spaces.GPU() for proper GPU allocation
5. ✅ Memory optimization - Added garbage collection and CUDA cache clearing

The error should be resolved and your Space should run on ZeroGPU infrastructure properly.