The Artificial Intelligence Controller Interface (AICI) lets you build Controllers that constrain and direct output of a Large Language Model (LLM) in real time. Controllers are flexible programs capable of implementing constrained decoding, dynamic editing of prompts and generated text, and coordinating execution across multiple, parallel generations. Controllers incorporate custom logic during the token-by-token decoding and maintain state during an LLM request. This allows diverse Controller strategies, from programmatic or query-based decoding to multi-agent conversations to execute efficiently in tight integration with the LLM itself.

The purpose of AICI is to make it easy to build and experiment with both existing and entirely new Controller strategies for improving LLM generations. By abstracting away implementation details of the underlying LLM inference and serving engine, AICI aims to simplify the development of Controllers, make it easier to write fast Controllers, and ease compatibility across LLM inference and serving engines.

AICI is designed for both local and cloud execution, including (eventually) multi-tenant LLM deployments. Controllers are implemented as light-weight WebAssembly (Wasm) modules which run on the same machine as the LLM inference engine, utilizing the CPU while the GPU is busy with token generation. AICI is one layer in the inference stack, and is designed to allow control libraries such as Guidance, LMQL, and others to run on top of it and gain both efficiency and performance improvements, as well as portability across LLM inference and serving engines.

AICI currently integrates with llama.cpp, HuggingFace Transformers, and rLLM (custom tch-based LLM inference engine), with vLLM in the works.

AICI is:

• Flexible: Controllers can be written in any language that can compile to Wasm (Rust, C, C++, ...), or be interpreted inside Wasm (Python, JavaScript, ...)
• Secure: Controllers are sandboxed and cannot access the filesystem, network, or any other resources
• Fast: Wasm modules are compiled to native code and run in parallel with the LLM inference engine, inducing only a minimal overhead to the generation process

AICI is a prototype, designed and built at Microsoft Research.

• Artificial Intelligence Controller Interface (AICI)
• QuickStart: Example Walkthrough
  • Development Environment Setup
  • Build and start rLLM server and AICI Runtime
  • Control AI output using AICI controllers
• Comprehensive Guide: Exploring Further
• Architecture
• Security
• Performance
• Flexibility
• Acknowledgements
• Contributing
• Trademarks

In this quickstart, we'll guide you through the following steps:

• Set up rLLM Server and AICI Runtime.
• Build and deploy a Controller.
• Use AICI to control LLM output, so you can customize a LLM to follow specific rules when generating text.

To compile AICI components, you need to set up your development environment for Rust. For this quickstart you also need Python 3.11 or later to create a controller.

If you're using devcontainer, you can skip to the next section.

Using the system package manager, install the necessary tools for building code in the repository, including git, cmake and ccache.

For instance in WSL / Ubuntu using apt:

sudo apt-get install --assume-yes --no-install-recommends \
    build-essential cmake ccache pkg-config libssl-dev libclang-dev clang llvm-dev git-lfs

or using Homebrew on macOS:

brew install git cmake ccache

Then install Rust, Rustup and Cargo, following the instructions provided here and here:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

After installation, verify that the rustup --version command is accessible by running it from the terminal. If the command isn't recognized, try opening a new terminal session.

Next install wasm32-wasi Rust component:

rustup target add wasm32-wasi

If you already had Rust installed, or are getting complaints from Cargo about outdated versions, run:

rustup update

Last, to work with Python controllers and scripts (like this tutorial), run this command to install the required packages:

pip install pytest pytest-forked ujson posix_ipc numpy requests

The rLLM server has two backends, one based on libtorch and CUDA (rllm-cuda), and the other based on llama.cpp (rllm-llamacpp).

The rllm-cuda backend only works with NVidia GPUs with compute capability 8.0 or later (A100 and later; RTX 30x0 and later) and requires a fiddly setup of libtorch -- it's strongly recommended to use the included devcontainer. While this guide focuses on the rllm-llamacpp backend, the build steps are the same for rllm-cuda, modulo the folder name.

After dev env setup above, clone the AICI repository and proceed with the next steps outlined below.

Use the following command to build and run aicirt and rllm-llamacpp:

cd rllm/rllm-llamacpp
./server.sh phi2

You can pass other model names as argument (run ./server.sh without arguments to see available models). You can also use a HuggingFace URL to .gguf file or a local path to a .gguf file. (For rllm-cuda use HuggingFace model id or path to folder).

./server.sh orca

You can find more details about rllm-llamacpp here.

The rLLM server provides a HTTP interface, utilized for configuration tasks and processing requests. You can also use this interface to promptly verify its status. For instance, if you open http://127.0.0.1:4242/v1/models, you should see:

{
  "object": "list",
  "data": [
    {
      "object": "model",
      "id": "TheBloke/phi-2-GGUF",
      "created": 946810800,
      "owned_by": "owner"
    }
  ]
}

confirming that the selected model is loaded.

AICI allows hosting custom logic, called Controllers, that initiate, terminate, and interact with LLMs token generation. Controllers take input arguments, process them, and return a result with logs, LLM tokens, and variables.

The repository includes some examples, in particular:

• jsctrl: a controller that accepts JavaScript code as input for execution. This code can interact with the model to generate text and tokens.
• pyctrl: a controller that accepts Python code as input for execution. This code can also interact with the model to generate text and tokens.

In this example we'll utilize pyctrl to manage token generation using a simple Python script. If you want, you can build and upload pyctrl, however by default the server will automatically download the latest release of pyctrl from GitHub.

In general, controllers require building and deployment, while scripts (Python or JavaScript) are sent with each request.

The following illustrates the relationship between the rLLM server, the AICI runtime, and the controller:

erDiagram
    Host    ||--|{ CPU : ""
    Host    ||--|{ GPU : ""
    
    CPU     ||--|| "rLLM Server" : execute
    CPU     ||--|{ "AICI Runtime" : execute

    "AICI Runtime" ||--|| "Controller" : instantiate

    GPU     ||--|{ "LLM token generation" : execute

Suppose we aim for a model to generate a list, adhering to a specific format and containing only five items.

Typically, achieving this involves prompt engineering, crafting the prompt precisely with clear instructions, such as:

What are the five most popular types of vehicles?
Return the result as a numbered list.
Do not add explanations, only the list.

The prompt would also vary depending on the model in use, given that each model tends to add explanations and understands instructions in different ways.

With AICI, we shift control back to code, and we can simplify the prompt to:

What are the most popular types of vehicles?

using code to:

1. Limit the list to 5 items
2. Prevent the model from adding some initial explanation
3. Format to a numbered list
4. Stop the model from adding some text after the list.

Let's create a list-of-five.py python file with the following content:

import pyaici.server as aici

# Force the model to generate a well formatted list of 5 items, e.g.
#   1. name 1
#   2. name 2
#   3. name 3
#   4. name 4
#   5. name 5
async def main():
    
    # This is the prompt we want to run.
    # Note how the prompt doesn't mention a number of vehicles or how to format the result.
    prompt = "What are the most popular types of vehicles?\n"

    # Tell the model to generate the prompt string, ie. let's start with the prompt "to complete"
    await aici.FixedTokens(prompt)

    # Store the current position in the token generation process
    marker = aici.Label()

    for i in range(1,6):
      # Tell the model to generate the list number
      await aici.FixedTokens(f"{i}.")

      # Wait for the model to generate a vehicle name and end with a new line
      await aici.gen_text(stop_at = "\n")

    await aici.FixedTokens("\n")

    # Store the tokens generated in a result variable
    aici.set_var("result", marker.text_since())

aici.start(main())

Running the script is not too different from sending a prompt. In this case, we're sending control logic and instructions all together.

To see the final result, execute the following command:

./aici.sh run list-of-five.py

Result:

Running with tagged AICI Controller: gh:microsoft/aici/pyctrl
[0]: FIXED 'What are the most popular types of vehicles?\n'
[0]: FIXED '1.'
[0]: GEN ' Cars\n'
[0]: FIXED '2.'
[0]: GEN ' Motorcycles\n'
[0]: FIXED '3.'
[0]: GEN ' Bicycles\n'
[0]: FIXED '4.'
[0]: GEN ' Trucks\n'
[0]: FIXED '5.'
[0]: GEN ' Boats\n'
[0]: FIXED '\n'
[DONE]
[Response] What are the most popular types of vehicles?
1. Cars
2. Motorcycles
3. Bicycles
4. Trucks
5. Boats

response saved to tmp/response.json
Usage: {'sampled_tokens': 16, 'ff_tokens': 37, 'cost': 69}
Timing: {'http_response': 0.05193686485290527, 'data0': 0.05199289321899414, 'first_token': 0.0658726692199707, 'last_token': 0.1784682273864746}
Tokens/sec: {'prompt': 861.0913072488067, 'sampling': 89.65181217019571}
Storage: {'result': '1. Cars\n2. Motorcycles\n3. Bicycles\n4. Trucks\n5. Boats\n\n'}

This repository contains a number of components, and which ones you need depends on your use case.

You can use an existing controller module. We provide PyCtrl and JsCtrl that let you script controllers using server-side Python and JavaScript, respectively. The pyaici package contains aici command line tool that lets you upload and run scripts with any controller (we also provide REST API definition for the curious).

🧑‍💻Python code samples for scripting PyCtrl and a JavaScript Hello World for JSCtrl

We anticipate libraries will be built on top of controllers. We provide an example in promptlib - a client-side Python library that generates interacts with DeclCtrl via the pyaici package.

🧑‍💻 Example notebook that uses PromptLib to interact with DeclCtrl.

The controllers can be run in a cloud or local AICI-enabled LLM inference engine. You can run the provided reference engine (rLLM) locally with either libtorch+CUDA or llama.cpp backend.

To develop a new controller, use a Rust starter project that shows usage of aici_abi library, which simplifies implementing the low-level AICI interface.

🧑‍💻Sample code for a minimal new controller to get you started

To add AICI support to a new LLM inference engine, you will need to implement LLM-side of the protocol that talks to AICI runtime.

Finally, you may want to modify any of the provided components - PRs are most welcome!

AICI abstracts LLM inference engine from the controller and vice-versa, as in the picture below. The rounded nodes are aspirational. Additional layers can be built on top - we provide promptlib, but we strongly believe that Guidance, LMQL, SGLang, Outlines, jsonformer, LMFE, etc. can also run on top of AICI (either with custom controllers or utilizing PyCtrl or JsCtrl).

graph TD
    PyCtrl -- AICI --> aicirt[AICI-runtime]
    JsCtrl -- AICI --> aicirt
    guidance([GuidanceCtrl]) -- AICI --> aicirt
    lmql([LMQL Ctrl]) -- AICI --> aicirt
    aicirt -- POSIX SHM --> rLLM
    aicirt -- POSIX SHM --> llama[llama.cpp]
    aicirt -- POSIX SHM --> pyaici
    pyaici -- Python --> vLLM(vLLM)
    pyaici -- Python --> hf[HF Transformers]

The pyaici package makes it easier to integrate AICI with Python-based LLM inference engines. Take a look at integration with HuggingFace Transformers, though note that it doesn't support forking (generation of multiple sequences in parallel). The vLLM REST server is currently out of date. Please use the rLLM-cuda or rLLM-llama.cpp for now.

• aicirt runs in a separate process, and can run under a different user than the LLM engine
• Wasm modules are sandboxed by Wasmtime
• Wasm only have access to aici_host_* functions, implemented in hostimpl.rs
• aicirt also exposes a partial WASI interface; however almost all the functions are no-op, except for fd_write which shims file descriptors 1 and 2 (stdout and stderr) to print debug messages
• each Wasm module runs in a separate process, helping with Spectre/Meltdown mitigation and allowing limits on CPU usage

In particular, Wasm modules cannot access the filesystem, network, or any other resources. They also cannot spin threads or access any timers (this is relevant for Spectre/Meltdown attacks).

Most of computation in AICI Controllers occurs on the CPU, in parallel with the logit generation on the GPU. The generation occurs in steps, where logits are generated in parallel for a new token for each sequence in a batch (typically between 1 and 50). This involves reading the whole model and KV caches for sequences in the batch from the GPU memory. For optimal batch throughput, the model and KV caches should utilize a major fraction of the GPU memory, and reading the whole memory takes about 40ms on A100 GPU (80GB).

Thus, each step of generation takes on the order of 20-50ms. With careful engineering, this is more than enough to compute the set of allowed tokens in Rust compiled to Wasm. These can be combined either natively in Rust, or via Python or JavaScript interpreters we provide.

For example, computing allowed token set in the 32000-strong vocabulary of Llama model takes:

• about 2.0ms for Yacc grammar of the C programming language
• about 0.3ms for a regular expression
• about 0.2ms for a substring constraint, from 4kB string

The above numbers are for a single sequence, however each sequence is processed in separate process, and thus if there is more cores than sequences (which is typical), they do not change. They also include overhead of calling into Python interpreter implemented in Wasm, and then back into Rust-generated Wasm code for the constraint itself. They are all well within the 20-50ms budget, so do not affect the generation time at all.

There is also some overhead in the critical path of sampling. It comes down to about 0.3ms per generation step when executing 10 sequences in parallel (this is irrespective of the constraint used). The overhead goes up to around 0.7ms for 40 sequences (though it has not been fully optimized yet).

WebAssembly is designed to have minimal overhead, compared to native code. In our experience, highly optimized Rust code is less than 2x slower when run in Wasmtime than native. This is 10-100x better than JavaScript or Python.

All measurements done on AMD EPYC 7V13 with nVidia A100 GPU with 80GB of VRAM.

The low-level interface that AICI runtime provides allows for:

• interaction with the LLM inference engine before, during, and after every generated token
• constraining decoding to a set of tokens
• backtracking KV-cache to a previous state
• fast-forwarding several tokens at a time (if they are known)
• forking generation into multiple branches
• communication between forks via shared variables
• utility functions for converting between tokens and byte strings

It can be utilized from any language that compiles to Wasm.

This repository provides a Rust library that makes it easy to implement controllers in Rust, and provides efficient implementations of specific constraints (regular expressions, yacc grammars, substrings). We also provide Python and JavaScript interpreters that allow to glue these constraints together. All of these can be easily extended.

• Flash Attention kernels are copied from flash-attention repo; see BSD LICENSE
• Paged Attention kernels are copied from vLLM repo; see Apache LICENSE
• OpenAI API definitions are copied and modified from candle-vllm; see MIT LICENSE
• cache_engine.rs, config.rs, and scheduler.rs are loosely based on vLLM
• llama.rs, phi.rs and logits.rs are based on candle-transformers
• specific Python library files are copied from RustPython (as we only use a subset of them)
• the example ANSI C grammar is based on https://www.lysator.liu.se/c/ANSI-C-grammar-y.html by Jeff Lee (from 1985)

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.