language-pipes 2.3.0

Peer-to-peer distributed inference for open-source language models

 _                                                   ____   _
| |                                                 |  __`\(_)                
| |     __ _  ___   ___  _   _  __ _  __ _  ___     | |__) | |_ __   ___  ___ 
| |    / _` |/ _ \ / _ `| | | |/ _` |/ _` |/ _ \    |  ___/| | '_ \ / _ \/ __|
| |___| (_| | | | | (_| | |_| | (_| | (_| |  __/    | |    | | |_) |  __/\__ \
|______\__,_|_| |_|\__, |\__,_|\__,_|\__, |\___|    |_|    |_| .__/ \___||___/
                    __/ |            __/ |                   | |              
                   |___/            |___/                    |_|      

Peer-to-peer distributed inference for open-source language models

Documentation & site: https://languagepipes.com

Language Pipes is an open-source distributed inference system built on the transformers library that splits large language model computation across multiple machines. By separating the model's text-handling components (embedding and output head) from its intermediate transformer layers, Language Pipes enables peer-to-peer inference.

Features

• OpenAI-compatible API
• Automatic model download by HuggingFace ID
• Interactive TUI for configuration, monitoring, and control
• Decentralized peer-to-peer network with optional AES encryption

---

How It Works

Language models process input through a sequence of transformer layers. Each layer performs matrix multiplications between learned weights and a hidden state tensor, passing the result to the next layer. Language Pipes distributes these layers across machines, splitting the memory cost across the network while keeping the text-handling components on the origin node.

The architecture provides architectural separation: layer models operate on continuous-valued tensors rather than discrete text while the end models keep text data on trusted systems. The privacy documentation provides a probabilistic threat model that quantifies the difficulty of known inversion attacks under various mitigation configurations.

Further reading:

• Architecture Overview: runtime components and inference flow
• Job Processor State Machine: how jobs traverse the distributed pipeline

Installation

Requires Python 3.10+. A GPU is optional — each node can run on CPU or CUDA, and a node only needs enough memory to hold the layers it hosts (memory scales with how many layers you assign it). For GPU support, install the appropriate PyTorch version for your CUDA configuration:
https://pytorch.org/get-started/locally/

See System requirements and network requirements for details, including what happens when a node drops mid-generation.

Install from pip:

pip install language-pipes

Quick Start

Launch the interactive TUI:

language-pipes

From the main menu, select New Configuration and give it a name to create a TOML config and open the dashboard (or Load Configuration to reopen one you've created before).

The dashboard is organized into tabs along the top: Home, Network, Models, Pipes, and Jobs. A fresh configuration has no node ID yet, so the only option on Home is Configure Network Server. Set a Node ID under Network > Configure, then return to Home and select Start Network Server. Once the network is running, the dashboard exposes the rest of setup: load models under Models > Layer Models / End Models, and configure and start the OpenAI-compatible API under Jobs > Server.

Configuration can also be edited directly as TOML files and run headlessly. See the CLI reference for details on running a saved configuration from the command line.

---

Two Node Example

This example distributes Qwen/Qwen3-1.7B across two computers. Node 1 hosts the End Model, so prompts and responses stay on Node 1, plus enough layers to fit in its memory. Node 2 hosts the remaining layers.

Node 1 (First Computer)

language-pipes

Select New Configuration and name it (e.g. node-1).

1. Network > Configure: set Node ID to node-1 and ensure Network IP is set to this machine's local IP address. Leave Network Key empty to disable encryption for this example. Peer Port defaults to 5000.
2. Back on Home, select Start Network Server.
3. Models > Installed: select Install New Model and enter Qwen/Qwen3-1.7B to download it.
4. Models > Layer Models: select Add Layer Model, choose Qwen/Qwen3-1.7B, a device (cpu or cuda:0), and a memory budget in GB (e.g. 2), then Save Model. Confirm to load it now.
5. Models > End Models: select Add End Model, choose Qwen/Qwen3-1.7B, and confirm to load it now.
6. Jobs > Server: ensure the Port is set to 8000 and select Start Server.

Node 2 (Second Computer)

language-pipes

Select New Configuration and name it (e.g. node-2).

1. Network > Configure: set Node ID to node-2. Under Bootstrap Nodes, add an entry with node-1's IP address and peer port (5000) so this node joins node-1's network.
2. Back on Home, select Start Network Server.
3. Models > Installed: install Qwen/Qwen3-1.7B as on Node 1.
4. Models > Layer Models: add Qwen/Qwen3-1.7B with a device and memory budget covering the remaining layers (e.g. 2 on cpu).

Once both nodes have loaded their layers, Pipes > Complete shows a completed pipe for Qwen/Qwen3-1.7B, and the model is ready for inference via node-1's Job Port.

Test the API

The model is accessible via the OpenAI-compatible API.

Example using the OpenAI Python library:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",  # node-1 IP address and Job Port
    api_key="not-needed"  # only required if api_keys is set in the config
)

response = client.chat.completions.create(
    model="Qwen/Qwen3-1.7B",
    max_completion_tokens=100,
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Write a haiku about distributed systems."}
    ]
)

print(response.choices[0].message.content)

Install the OpenAI library with: pip install openai

See the OpenAI-compatible API documentation for the full endpoint reference and sampling parameter descriptions.

Supported Models

Language Pipes supports Qwen3 (dense and MoE), Phi, Meta Llama 2/3, Gemma 3, Gemma 4, and Ministral 3. See the supported-models list for a more complete list.

Roadmap

• Additional model architectures
• 4 bit quantization
• KV Cache quantization
• GGUF format support (currently requires safetensors)
• Failover/rerouting when a layer node drops, and integrity guarantees for layer computation

See the full roadmap for details and open research directions.

Contributing

Contributions are welcome! Bug reports, documentation fixes, and especially new model architectures. See CONTRIBUTING.md for dev setup, how to run the tests, and PR expectations, and the Code of Conduct. Questions and ideas are welcome in GitHub Discussions.

Packages

Language Pipes is built on two components that are also published as standalone PyPI packages, maintained from this repo under packages/:

• llm-layer-collector — load individual transformer components (embedding, decoder layers, norm, head) from sharded HuggingFace checkpoints and run per-architecture computation. Docs
• distributed-state-network — an encrypted peer-to-peer state-sharing network over HTTP. Docs

Language Pipes pins exact versions of both, so pip install language-pipes always pulls the matching releases — you never have to update them yourself.

Dependencies

• pytorch
• transformers

Documentation

The docs are published as a website at https://erinclemmer.github.io/language-pipes (built from this folder by website/). The Markdown source of truth lives in documentation/:

• CLI Reference
• Privacy Protection
• Configuration Manual
• Architecture Overview
• OpenAI-Compatible API
• Job Processor State Machine
• Request For Model Protocol
• Distributed State Network
• LLM Layer Collector
• Release Notes