IPFS Accelerate Python

Enterprise-grade hardware-accelerated machine learning inference with IPFS network-based distribution

---

📋 Table of Contents

• Overview
• Installation
• Quick Start
• MCP++ Server
• Architecture
• Supported Hardware
• Supported Models
• Documentation
• IPFS & Distributed Features
• Performance & Optimization
• Troubleshooting
• Testing & Quality
• Contributing
• License

---

🚀 Overview

IPFS Accelerate Python combines cutting-edge hardware acceleration, distributed computing, and IPFS network integration to deliver blazing-fast machine learning inference across multiple platforms and devices - from data centers to browsers.

⚡ Key Highlights

• 🔥 8+ Hardware Platforms - CPU, CUDA, ROCm, OpenVINO, Apple MPS, WebNN, WebGPU, Qualcomm
• 🌐 Distributed by Design - IPFS content addressing, P2P inference, global caching
• 🤖 300+ Models - Full HuggingFace compatibility + custom architectures
• 🧠 Canonical MCP++ Server - Unified ipfs_accelerate_py.mcp_server runtime is now the default startup path
• 🌍 Browser-Native - WebNN & WebGPU for client-side acceleration
• 📊 Production Ready - Real-time monitoring, enterprise security, compliance validation
• ⚡ High Performance - Intelligent caching, batch processing, model optimization

---

📦 Installation

Quick Start (5 minutes)

# 1. Create virtual environment
python3 -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

# 2. Install IPFS Accelerate
pip install -U pip setuptools wheel
pip install ipfs-accelerate-py

# 3. Verify installation
python -c "from ipfs_accelerate_py import IPFSAccelerator; print('✅ Ready!')"

NVIDIA CUDA (PyTorch)

By default, pip may install a CPU-only PyTorch wheel from PyPI (e.g. torch==...+cpu) because the CUDA-enabled wheels are published on PyTorch's wheel indexes.

If you have an NVIDIA GPU and want to ensure CUDA is available in PyTorch, install PyTorch from the CUDA wheel index:

python -m pip install -U pip
python -m pip install --upgrade --force-reinstall -r install/requirements_torch_cu124.txt

python -c "import torch; print('torch=', torch.__version__); print('cuda_available=', torch.cuda.is_available()); print('torch_cuda=', torch.version.cuda)"

If you're on an NVIDIA GB10 / DGX Spark-class system (CUDA capability 12.1, CUDA 13.0), stable builds may warn that your GPU is unsupported. In that case, use the CUDA 13.0 nightly wheels:

./scripts/install_torch_cuda_cu130_nightly.sh

If you're installing from source/editable mode, you can also run:

python -m pip install -e . --no-deps
python -m pip install --upgrade --force-reinstall -r install/requirements_torch_cu124.txt
python -m pip install -r requirements.txt

Installation Profiles

Choose the profile that matches your needs:

Profile	Use Case	Installation
Core	Basic inference	pip install ipfs-accelerate-py
Full	Models + API server	pip install ipfs-accelerate-py[full]
MCP	MCP server extras	pip install ipfs-accelerate-py[mcp]
Dev	Development setup	pip install -e .

📚 Detailed instructions: Installation Guide | Troubleshooting | Getting Started

---

🎯 Quick Start

Python API

from ipfs_accelerate_py import IPFSAccelerator

# Initialize with automatic hardware detection
accelerator = IPFSAccelerator()

# Load any HuggingFace model
model = accelerator.load_model("bert-base-uncased")

# Run inference (automatically optimized for your hardware)
result = model.inference("Hello, world!")
print(result)

Command Line Interface

# Start the default MCP++ server for automation
ipfs-accelerate mcp start

# Run the canonical FastAPI MCP service directly
python -m ipfs_accelerate_py.mcp_server.fastapi_service

# Run the direct MCP server CLI with p2p/task options
python -m ipfs_accelerate_py.mcp.cli --host 0.0.0.0 --port 9000

# Run inference directly
ipfs-accelerate inference generate \
  --model bert-base-uncased \
  --input "Hello, world!"

# List available models and hardware
ipfs-accelerate models list
ipfs-accelerate hardware status

# Start GitHub Actions autoscaler
ipfs-accelerate github autoscaler

Remote libp2p task pickup (ipfs_datasets_py)

If you want a remote machine running the ipfs_accelerate_py MCP server to also pick up libp2p task submissions coming from ipfs_datasets_py, you can start the MCP server CLI with the built-in P2P task worker:

# Remote machine (runs MCP + worker + libp2p TaskQueue service)
python -m ipfs_accelerate_py.mcp.cli \
  --host 0.0.0.0 --port 9000 \
  --p2p-task-worker --p2p-service --p2p-listen-port 9710 \
  --p2p-queue ~/.cache/ipfs_datasets_py/task_queue.duckdb

# Optional (off-host clients): set the public IP that will be embedded in the announced multiaddr
export IPFS_DATASETS_PY_TASK_P2P_PUBLIC_IP="YOUR_PUBLIC_IP"

By default, the libp2p TaskQueue service writes an announce file into your XDG cache dir and clients will try to use it automatically:

• Default announce file: ~/.cache/ipfs_accelerate_py/task_p2p_announce.json
• Disable announce file (opt-out): IPFS_ACCELERATE_PY_TASK_P2P_ANNOUNCE_FILE=0 (or IPFS_DATASETS_PY_TASK_P2P_ANNOUNCE_FILE=0)

If your client machine can read that announce file (same host/user, or a shared filesystem path you set via IPFS_ACCELERATE_PY_TASK_P2P_ANNOUNCE_FILE / IPFS_DATASETS_PY_TASK_P2P_ANNOUNCE_FILE), you do not need to set any remote multiaddr env vars.

Otherwise, the process also prints a multiaddr=... line. On the client machine, set:

export IPFS_DATASETS_PY_TASK_P2P_REMOTE_MULTIADDR="/ip4/.../tcp/9710/p2p/..."

Notes:

• This mode requires ipfs_datasets_py to be installed on the remote machine (and libp2p installed via ipfs_datasets_py[p2p]).

Real-World Examples

Example	Description	Complexity
Basic Usage	Simple inference with BERT	Beginner
Hardware Selection	Choose specific accelerator	Intermediate
Distributed Inference	P2P model sharing	Advanced
Browser Integration	WebNN/WebGPU in browsers	Advanced

📖 More examples: examples/ | Quick Start Guide

---

🧠 MCP++ Server

The MCP server in this repository has completed its unification cutover.

• Canonical runtime: ipfs_accelerate_py/mcp_server
• Compatibility facade: ipfs_accelerate_py/mcp
• Current default: create_mcp_server() and the main MCP startup paths now select the unified runtime by default
• Cutover status: approved and frozen with a focused release-candidate matrix of 120 passed

Current entrypoints

Entry point	Best for	Notes
ipfs-accelerate mcp start	End-user server startup	Main product CLI for MCP server management and dashboard workflows
python -m ipfs_accelerate_py.mcp.cli	Direct server/process control	Starts the MCP server and can also host TaskQueue/libp2p worker services
python -m ipfs_accelerate_py.mcp_server.fastapi_service	Standalone HTTP/FastAPI hosting	Reads IPFS_MCP_* env vars and mounts the MCP app at /mcp by default
from ipfs_accelerate_py.mcp_server import create_server	Programmatic embedding	Stable import target for the canonical runtime package

Supported MCP++ profile chapters

The unified runtime currently advertises these additive MCP++ profiles:

• mcp++/profile-a-idl
• mcp++/profile-b-cid-artifacts
• mcp++/profile-c-ucan
• mcp++/profile-d-temporal-policy
• mcp++/profile-e-mcp-p2p

Unified control-plane features

• Meta-tools: tools_list_categories, tools_list_tools, tools_get_schema, tools_dispatch, tools_runtime_metrics
• Migrated native categories: ipfs, workflow, p2p
• Security and governance: UCAN validation, temporal/deontic policy evaluation, policy audit logging, secrets vault support, and risk scoring/frontier execution
• Observability: runtime metrics, audit-to-metrics bridging, OpenTelemetry hooks, and Prometheus exporter support
• Transport coverage: compatibility-tested process helpers, FastAPI mounting, and MCP+p2p handler parity with mixed-version negotiation hardening

Cutover and rollback controls

These controls remain available for validation and operational rollback:

• IPFS_MCP_FORCE_LEGACY_ROLLBACK=1 — force the compatibility facade to stay on the legacy wrapper
• IPFS_MCP_UNIFIED_CUTOVER_DRY_RUN=1 — validate the unified startup path while keeping legacy runtime behavior active
• IPFS_MCP_ENABLE_UNIFIED_BRIDGE=1 — explicitly request the unified bridge on compatibility-facade paths

Recommended documentation

• Canonical MCP server README
• MCP Cutover Checklist
• MCP Server Unification Plan
• MCP++ Conformance Checklist
• MCP++ Spec Gap Matrix

---

🏗️ Architecture

IPFS Accelerate Python is built on a modular, enterprise-grade architecture:

┌─────────────────────────────────────────────────────────┐
│                   Application Layer                     │
│  Python API • CLI • MCP Server • Web Dashboard          │
└────────────────────┬────────────────────────────────────┘
                     │
┌────────────────────┴────────────────────────────────────┐
│              Hardware Abstraction Layer                 │
│  Unified interface across 8+ hardware platforms         │
└────────────────────┬────────────────────────────────────┘
                     │
┌────────────────────┴────────────────────────────────────┐
│                Inference Backends                       │
│  CPU • CUDA • ROCm • MPS • OpenVINO • WebNN • WebGPU    │
└────────────────────┬────────────────────────────────────┘
                     │
┌────────────────────┴────────────────────────────────────┐
│              IPFS Network Layer                         │
│  Content addressing • P2P • Distributed caching         │
└─────────────────────────────────────────────────────────┘

Core Components

• Hardware Abstraction: Unified API across 8+ platforms with automatic selection
• IPFS Integration: Content-addressed storage, P2P distribution, intelligent caching
• Performance Modeling: ML-powered optimization and resource management
• MCP Server: Canonical ipfs_accelerate_py.mcp_server MCP++ runtime with compatibility facade and cutover controls
• Monitoring: Real-time metrics, profiling, and analytics

📐 Detailed architecture: docs/architecture/overview.md | CI/CD

---

🔧 Supported Hardware

Run anywhere - from powerful servers to edge devices and browsers:

Platform	Status	Acceleration	Requirements	Performance
CPU (x86/ARM)	✅	SIMD, AVX	Any	Good
NVIDIA CUDA	✅	GPU + TensorRT	CUDA 11.8+	Excellent
AMD ROCm	✅	GPU + HIP	ROCm 5.0+	Excellent
Apple MPS	✅	Metal	M1/M2/M3	Excellent
Intel OpenVINO	✅	CPU/GPU	Intel HW	Very Good
WebNN	✅	Browser NPU	Chrome, Edge	Good
WebGPU	✅	Browser GPU	Modern browsers	Very Good
Qualcomm	✅	Mobile DSP	Snapdragon	Good

Hardware Selection

The framework automatically detects and selects the best available hardware:

# Automatic (recommended)
accelerator = IPFSAccelerator()  # Uses best available

# Manual selection
accelerator = IPFSAccelerator(device="cuda")  # Force CUDA
accelerator = IPFSAccelerator(device="mps")   # Force Apple MPS

⚙️ Hardware guides: Hardware Optimization | Platform Support

---

🤖 Supported Models

Pre-trained Models (300+)

Category	Models	Status
Text	BERT, RoBERTa, DistilBERT, ALBERT, GPT-2/Neo/J, T5, BART, Pegasus, Sentence Transformers	✅
Vision	ViT, DeiT, BEiT, ResNet, EfficientNet, DETR, YOLO	✅
Audio	Whisper, Wav2Vec2, WavLM, Audio Transformers	✅
Multimodal	CLIP, BLIP, LLaVA	✅
Custom	PyTorch models, ONNX, TensorFlow (converted)	✅

Model Loading

# From HuggingFace Hub
model = accelerator.load_model("bert-base-uncased")

# From IPFS (content-addressed)
model = accelerator.load_model("ipfs://QmXxxx...")

# Local model
model = accelerator.load_model("./my_model/")

# With specific hardware
model = accelerator.load_model("gpt2", device="cuda")

🤖 Full model list: Supported Models | Custom Models Guide

---

📚 Documentation

📖 Essential Guides

Guide	Description	Audience
Getting Started	Complete beginner tutorial	Everyone
Quick Start	Get running in 5 minutes	Everyone
Installation	Detailed setup instructions	Users
FAQ	Common questions & answers	Everyone
API Reference	Complete API documentation	Developers
Architecture	System design & components	Architects
Hardware Optimization	Platform-specific tuning	Engineers
Testing Guide	Testing & benchmarking	QA/DevOps

🎯 Specialized Topics

Topic	Resources
IPFS & P2P	IPFS Integration • P2P Networking
GitHub Actions	Autoscaler • CI/CD
Docker & K8s	Container Guide • Deployment
MCP Server	Canonical MCP Server README • MCP Setup • Protocol Docs • Cutover Checklist
Browser Support	WebNN/WebGPU • Examples

📊 Documentation Quality

Our documentation has been professionally audited (January 2026):

• ✅ 200+ files covering all features
• ✅ 93/100 quality score (Excellent)
• ✅ Comprehensive - From beginner to expert
• ✅ Well-organized - Clear structure and navigation
• ✅ Verified - All examples tested and working

📋 Documentation Hub: docs/ | Full Index | Audit Report

---

🌐 IPFS & Distributed Features

Why IPFS?

IPFS integration provides enterprise-grade distributed computing:

• 🔐 Content Addressing - Cryptographically secure, immutable model distribution
• 🌍 Global Network - Automatic peer discovery and geographic optimization
• ⚡ Intelligent Caching - Multi-level LRU caching across the network
• 🔄 Load Balancing - Automatic distribution across available peers
• 🛡️ Fault Tolerance - Robust error handling and fallback mechanisms

IPFS Backend Router (New! ⭐)

The IPFS Backend Router provides a flexible, pluggable backend system with automatic fallback:

Backend Preference Order:

1. ipfs_kit_py - Full distributed storage (preferred)
2. HuggingFace Cache - Local storage with IPFS addressing
3. Kubo CLI - Standard IPFS daemon

from ipfs_accelerate_py import ipfs_backend_router

# Store model weights to IPFS
cid = ipfs_backend_router.add_path("/path/to/model", pin=True)
print(f"Model CID: {cid}")

# Retrieve from anywhere
ipfs_backend_router.get_to_path(cid, output_path="/cache/model")

Configuration:

# Prefer ipfs_kit_py (default)
export ENABLE_IPFS_KIT=true

# Use HF cache only (good for CI/CD)
export IPFS_BACKEND=hf_cache

# Force Kubo CLI
export IPFS_BACKEND=kubo

📚 Full documentation: IPFS Backend Router Guide

Distributed Inference

# Enable P2P inference
accelerator = IPFSAccelerator(enable_p2p=True)

# Model is automatically shared across peers
model = accelerator.load_model("bert-base-uncased")

# Inference uses best available peer
result = model.inference("Distributed AI!")

Advanced Features

Feature	Description	Status
P2P Workflow Scheduler	Distributed task execution with merkle clocks	✅
GitHub Actions Cache	Distributed cache for CI/CD	✅
Autoscaler	Dynamic runner provisioning	✅
MCP Server	Model Context Protocol (14+ tools)	✅

🌐 Learn more: IPFS Guide | P2P Architecture | Network Setup

---

🧪 Testing & Quality

# Run all tests
pytest

# Run specific test suite
pytest test/test_inference.py

# Run with coverage report
pytest --cov=ipfs_accelerate_py --cov-report=html

# Run benchmarks
python data/benchmarks/run_benchmarks.py

Quality Metrics

Metric	Status	Details
Test Coverage	✅	Comprehensive test suite
Documentation	✅ 93/100	Audit Report
Code Quality	✅	Linted, type-checked
Security	✅	Regular vulnerability scans
Performance	✅	Benchmarked across platforms

🧪 Testing guide: docs/guides/testing/TESTING_README.md | CI/CD Setup

---

⚡ Performance & Optimization

Benchmarks

Hardware	Model	Throughput	Latency
NVIDIA RTX 3090	BERT-base	~2000 samples/sec	<1ms
Apple M2 Max	BERT-base	~800 samples/sec	2-3ms
Intel i9 (CPU)	BERT-base	~100 samples/sec	10-15ms
WebGPU (Browser)	BERT-base	~50 samples/sec	20-30ms

Optimization Tips

# Enable mixed precision for 2x speedup
accelerator = IPFSAccelerator(precision="fp16")

# Use batch processing for better throughput
results = model.batch_inference(inputs, batch_size=32)

# Enable model quantization for 4x memory reduction
model = accelerator.load_model("bert-base-uncased", quantize=True)

# Use intelligent caching for repeated queries
accelerator = IPFSAccelerator(enable_cache=True)

📊 Performance guide: Hardware Optimization | Benchmarking

---

🔧 Troubleshooting

Common Issues

Issue	Solution
Import errors	pip install --upgrade ipfs-accelerate-py
CUDA not found	Install CUDA Toolkit 11.8+
Slow inference	Check hardware selection, enable caching
Memory errors	Use quantization, reduce batch size
Connection issues	Check IPFS daemon, firewall settings

Quick Fixes

# Verify installation
python -c "import ipfs_accelerate_py; print(ipfs_accelerate_py.__version__)"

# Check hardware detection
ipfs-accelerate hardware status

# Test basic inference
ipfs-accelerate inference test

# View logs
ipfs-accelerate logs --tail 100

🆘 Get help: Troubleshooting Guide | FAQ | GitHub Issues

---

🤝 Contributing

We welcome contributions! Here's how to get started:

Quick Contribution Guide

1. Fork & Clone: Get your own copy of the repository
2. Create Branch: git checkout -b feature/your-feature
3. Make Changes: Follow our coding standards
4. Run Tests: pytest to ensure everything works
5. Submit PR: Open a pull request with clear description

Areas We Need Help

• 🐛 Bug Reports - Found an issue? Let us know!
• 📚 Documentation - Help improve guides and examples
• 🧪 Testing - Add tests for edge cases
• 🌍 Translations - Translate docs to other languages
• 💡 Features - Suggest or implement new features

Community & Guidelines

• 💬 GitHub Discussions - Ask questions, share ideas
• 🐛 Issue Tracker - Report bugs, request features
• 🔐 Security Policy - Report security vulnerabilities
• 📧 Email: starworks5@gmail.com

📖 Full guides: CONTRIBUTING.md | Code of Conduct | Security Policy

---

📄 License

This project is licensed under the GNU Affero General Public License v3.0 or later (AGPLv3+).

What this means:

• ✅ Free to use, modify, and distribute
• ✅ Commercial use allowed
• ✅ Patent protection included
• ⚠️ Source code must be disclosed for network services
• ⚠️ Modifications must use same license

📋 Details: LICENSE | AGPL FAQ

---

🙏 Acknowledgments

Built with amazing open source technologies:

• HuggingFace Transformers - ML model ecosystem
• IPFS - Distributed file system
• PyTorch - Deep learning framework
• FastAPI - Modern web framework

Special thanks to all contributors who make this project possible! 🌟

Project Information

• 📋 Changelog - Version history and release notes
• 🔐 Security Policy - Security reporting and best practices
• 🤝 Contributing Guide - How to contribute
• 📄 License - AGPLv3+ license details

---

🌟 Show Your Support

If you find this project useful:

• ⭐ Star this repository on GitHub
• 📢 Share with your network
• 🐛 Report issues to help improve it
• 💡 Contribute features or fixes
• 📝 Write about your experience

---

Made with ❤️ by Benjamin Barber and contributors

🏠 Homepage • 📚 Documentation • 🐛 Issues • 💬 Discussions