DEPLOYMENT.md

InfiniteTalk - Deployment Guide

Prerequisites

1. HuggingFace Account: Sign up at https://huggingface.co
2. Git & Git LFS: Install from https://git-scm.com
3. HuggingFace CLI (optional but recommended):

   pip install huggingface_hub
   huggingface-cli login
   

Deployment Steps

Option 1: Web UI (Easiest)

1. Create New Space

   • Go to https://huggingface.co/new-space
   • Space name: infinitetalk (or your choice)
   • License: apache-2.0
   • SDK: Gradio
   • Hardware: ZeroGPU (free tier available!)
   • Click "Create Space"

2. Upload Files

   • Click "Files" tab in your new Space
   • Upload all files from this directory:
     • README.md (with YAML metadata)
     • app.py
     • requirements.txt
     • packages.txt
     • .gitignore
     • src/ folder
     • wan/ folder
     • utils/ folder
     • assets/ folder (optional)
     • examples/ folder (optional)
     • LICENSE.txt

3. Wait for Build

   • Space will automatically build
   • First build takes 5-10 minutes (installing dependencies)
   • Check "Logs" tab for build progress
   • Watch for any error messages

4. Test Your Space

   • Once built, the Space will show "Running"
   • First generation will download models (~2-3 minutes)
   • Try with example images/audio

Option 2: Git (Advanced)

1. Clone Your Space

   git clone https://huggingface.co/spaces/YOUR_USERNAME/YOUR_SPACE_NAME
   cd YOUR_SPACE_NAME
   

2. Copy Files

   # From your local infinitetalk-hf-space directory
   cp -r /path/to/infinitetalk-hf-space/* .
   

3. Commit and Push

   git add .
   git commit -m "Initial InfiniteTalk Space deployment"
   git push
   

4. Monitor Build

   • Go to your Space URL
   • Check "Logs" for build progress

Option 3: CLI Upload

# From this directory
huggingface-cli upload YOUR_USERNAME/YOUR_SPACE_NAME . --repo-type=space

Troubleshooting

Build Fails with Flash-Attn Error

Symptom: flash-attn compilation fails

Solutions:

1. Try adding to requirements.txt:

   flash-attn==2.7.4.post1 --no-build-isolation
   

2. Or use Dockerfile approach (create Dockerfile):

   FROM nvidia/cuda:12.1.0-devel-ubuntu22.04
   
   RUN apt-get update && apt-get install -y \
       python3.10 python3-pip git ffmpeg build-essential libsndfile1
   
   WORKDIR /app
   
   # Install PyTorch first
   RUN pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1
   
   # Install flash-attn with pre-built wheels
   RUN pip install flash-attn==2.7.4.post1 --no-build-isolation
   
   # Copy and install requirements
   COPY requirements.txt .
   RUN pip install -r requirements.txt
   
   # Copy application
   COPY . .
   
   CMD ["python3", "app.py"]
   

Models Not Downloading

Symptom: "Model download failed" error

Solutions:

1. Check HuggingFace is not down: https://status.huggingface.co
2. Add HF_TOKEN secret in Space settings (for private models)
3. Check model repository IDs in utils/model_loader.py

Out of Memory (OOM) Errors

Symptom: "CUDA out of memory"

Solutions:

1. Reduce resolution (use 480p instead of 720p)
2. Reduce diffusion steps (try 30 instead of 40)
3. Process shorter videos
4. Check utils/gpu_manager.py settings

Space Stuck in "Building"

Symptom: Build takes >15 minutes

Solutions:

1. Check "Logs" tab for errors
2. Flash-attn compilation can take 10+ minutes
3. If timeout, try Dockerfile approach
4. Consider pre-built flash-attn wheels

ZeroGPU Quota Exceeded

Symptom: "GPU quota exceeded"

Solutions:

1. Free Tier: Wait for quota to refill (1 ZeroGPU second = 30 real seconds)
2. Upgrade to PRO: $9/month for 8× quota
3. Apply for Grant: Community GPU Grant for innovative projects
4. Optimize generation time (reduce steps, use 480p)

Post-Deployment

Monitor Usage

• Check "Logs" tab regularly
• Monitor GPU quota in Space settings
• Watch for user error reports in Community tab

Update Space

# Make changes locally
git add .
git commit -m "Update: [description]"
git push

Space will automatically rebuild on push.

Add Examples

Upload example images and audio to examples/ folder to help users get started quickly.

Enable Discussions

In Space settings, enable "Discussions" to get user feedback.

Apply for Community GPU Grant

If your Space is popular and useful:

1. Go to Space Settings
2. Click "Apply for community GPU grant"
3. Explain your project's value to the community

Hardware Options

Free ZeroGPU

• Cost: FREE
• Limits: 300s per session, 600s max quota
• Best for: Testing, light usage, demos
• GPU: H200 with 70GB VRAM

PRO ZeroGPU

• Cost: $9/month
• Benefits: 8× quota, priority queue, 10 Spaces
• Best for: Regular usage, public demos

Dedicated GPU (Paid)

• T4 (16GB): $0.60/hour - Too small for InfiniteTalk
• A10G (24GB): $1.05/hour - Minimum viable
• A100 (40GB): $3.00/hour - Overkill but works
• Best for: Private, dedicated instances

Performance Expectations

First Generation

• Model download: 2-3 minutes
• Generation (10s video, 480p): 40 seconds
• Total: ~3-4 minutes

Subsequent Generations

• Generation (10s video, 480p): 35-40 seconds
• Generation (10s video, 720p): 60-70 seconds

Free Tier Usage

• ~3-5 generations per quota period (600s ZeroGPU)
• Quota refills gradually (1 ZeroGPU second per 30 real seconds)

Support

• Issues: File at https://github.com/MeiGen-AI/InfiniteTalk/issues
• Discussions: Use Space's Community tab
• HF Forums: https://discuss.huggingface.co

Success Checklist

• Space builds without errors
• Models download successfully on first run
• Example image-to-video generation works
• Example video dubbing works
• No OOM errors with 480p
• GPU memory is cleaned up between runs
• Gradio UI is responsive
• Examples are loaded and working
• README displays correctly
• Space doesn't crash after multiple uses

Good luck with your deployment! 🚀