MinMochi

Minimalist API Server for Mochi Text-to-Video Generation

🚀 Overview

MinMochi serves the Genmo Mochi text-to-video model as a production-ready API. Generate high-quality videos from text prompts with minimal setup.

🛠️ System Requirements

• 🐍 Python 3.10+
• 🎮 GPU Requirements:
  • Recommended: NVIDIA A100 or H100
  • Suitable: NVIDIA A6000 or A40
• ☁️ Active AWS account
• 🐳 Docker

📦 Installation

# Get the code
git clone https://github.com/VikramxD/Minimochi
cd minimochi

# Set up environment
pip install uv
uv venv .venv
uv pip install -r requirements.txt
uv pip install -e . --no-build-isolation

⚙️ Configuration

MinMochi uses Pydantic settings for configuration management. The configuration is split into three main modules:

1. Mochi Settings (mochi_settings.py)

# Default settings, can be overridden with MOCHI_ prefixed env variables
model_name = "Genmo-Mochi"
transformer_path = "imnotednamode/mochi-1-preview-mix-nf4"
pipeline_path = "VikramxD/mochi-diffuser-bf16"
dtype = torch.bfloat16
device = "cuda"

# Optimization Settings
enable_vae_tiling = True
enable_model_cpu_offload = True
enable_attention_slicing = False

# Video Generation Settings
num_inference_steps = 20
guidance_scale = 7.5
height = 480
width = 848
num_frames = 150
fps = 10

2. AWS Settings (aws_settings.py)

# Override with environment variables
AWS_ACCESS_KEY_ID = ""
AWS_SECRET_ACCESS_KEY = ""
AWS_REGION = "ap-south-1"
AWS_BUCKET_NAME = "diffusion-model-bucket"

3. Model Weights Settings (mochi_weights.py)

output_dir = Path("weights")
repo_id = "genmo/mochi-1-preview"
model_file = "dit.safetensors"
decoder_file = "decoder.safetensors"
encoder_file = "encoder.safetensors"
dtype = "bf16"  # Options: "fp16", "bf16"

🎬 Usage

Launch Server

python src/api/mochi_serve.py

Generate Videos

import requests
import json

url = "http://localhost:8000/api/v1/video/mochi"
payload = {
    "prompt": "A beautiful sunset over the mountains",
    "negative_prompt": "",
    "num_inference_steps": 100,
    "guidance_scale": 7.5,
    "height": 480,
    "width": 848,
    "num_frames": 150,
    "fps": 10
}

response = requests.post(url, json=[payload])
print(response.json())

📊 Monitoring

Metrics

Prometheus metrics available at /metrics:

• Request processing time
• GPU memory usage
• Inference time

Logging

• Structured logging with loguru
• Log rotation at 100MB
• 1-week retention period
• Logs stored in logs/api.log

🎛️ GPU Memory Requirements

Resolution	Frames	Min GPU Memory
480x480	60	16GB
576x576	60	20GB
768x768	60	24GB

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Genmo.ai for the original Mochi model
• Hugging Face Diffusers
• LitServe - API framework

---

Report Bug • Request Feature