node-red-flow-stablelm-3b-4e1t

Node-RED Flows for the StableLM-3B-4E1T AI model

This repository contains a function node for Node-RED which can be used to run the Stability AI StableLM-3B-4E1T model using llama.cpp within a Node-RED flow. Inference is done on the CPU (without requiring any special harware) and still completes within a few seconds on a reasonably powerful computer.

Additionally, this repo also contains function nodes to tokenize a prompt or to calculate embeddings based on the StableLM-3B-4E1T model.

Having the inference, tokenization and embedding calculation as a self-contained function node gives you the possibility to create your own user interface or even use it as part of an autonomous agent.

Nota bene: these flows do not contain the actual model. You will have to download your own copy directly from Huggingface (use file stablelm-3b-4e1t-Q8_K.bin).

Just a small note: if you like this work and plan to use it, consider "starring" this repository (you will find the "Star" button on the top right of this page), so that I know which of my repositories to take most care of.

Installation

This section shows you how to install Node.js, Node-RED, LLaMA.cpp and the flows from this repository - feel free to skip the steps for those components you already installed before.

Node.js

"Node.js is a cross-platform, open-source server environment that can run on Windows, Linux, Unix, macOS, and more. Node.js is a back-end JavaScript runtime environment, runs on the V8 JavaScript engine, and executes JavaScript code outside a web browser." (according to Wikipedia)

Start by installing Node.js as described on their web page.

Node-RED

"Node-RED is a flow-based, low-code development tool for visual programming developed originally by IBM..." (according to Wikipedia)

If not already done, install Node-RED as described on their "Get Started" page.

LLaMA.cpp

LLaMA.cpp is a port of Facebook's LLaMA model in C/C++ (don't be fooled by the statement that the "main goal of llama.cpp is to run the LLaMA model using 4-bit integer quantization on a MacBook" - LLaMA.cpp runs just as well under Windows and Linux, and with different quantizations)

Note: please use my own fork of the original LLaMA.cpp as the code in there raises the context limit and contains additional functions for tokenization - however, if you are fine with a context length of up to 4096 tokens and don't want to tokenize any prompts, the original code will work as fine

Get the code as follows:

git https://github.com/rozek/llama.cpp
cd llama.cpp

Then continue as described in the LLaMA.cpp docs

Afterwards, rename the executables

• main to llama,
• tokenization to llama-tokens and
• embedding to llama-embeddings

and copy them into the Node-RED "User Directory" (by default, this folder is located at $HOME/.node-red).

StableLM-3B-4E1T Flows

If not already done, download the 8-bit quantization of the StableLM-3B-4E1T model into the same folder that already contains your executables.

Nota bene: right now, the flows from this repository support the given model file only - if you prefer another one, you may simply change the model file name in the function nodes for text completion, tokenization and embeddings calculation.

Now import the desired nodes and flows - if you want them all, just import file StableLM-3B-4E1T-Flows.json.

If you are new to Node-RED, just follow the instructions from their docs.

Configuration

By default, all StableLM-3B-4E1T function nodes expect their executables (and the model itself) in the folder $HOME/.node-red (where $HOME is the user's home directory)

If you prefer to store everything in a different folder, simply edit the "configure Settings" node from StableLM-3B-4E1T-Flows.json and set globals.UserDir to the fully qualified path of the directory you chose.

Function Node Usage

All function nodes expect their parameters as properties of the msg object. The prompt itself (or the input text to tokenize or calculate embeddings from) is expected in msg.payload and will later be replaced by the function result.

All properties (except prompt or input text) are optional. If given, they should be strings (even if they contain numbers), this makes it simpler to extract them from an HTTP request.

Text Completion Node

Text completion supports the following properties:

• payload - this is the actual prompt
• seed - seed value for the internal pseudo random number generator (integer, default: -1, use random seed for <= 0)
• threads - number of threads to use during computation (integer ≧ 1, default: 4)
• context - size of the prompt context (0...4096, default: 512)
• keep - number of tokens to keep from the initial prompt (integer ≧ -1, default: 0, -1 = all)
• predict - number of tokens to predict (integer ≧ -1, default: 128, -1 = infinity)
• topk - top-k sampling limit (integer ≧ 1, default: 40)
• topp - top-p sampling limit (0.0...1.0, default: 0.9)
• temperature - temperature (0.0...2.0, default: 0.8)
• batches - batch size for prompt processing (integer ≧ 1, default: 8)
• grammar - GBNF grammar to restrict text completion output

Tokenization Node

Tokenization supports the following properties:

• payload - this is the actual input text
• threads - number of threads to use during computation (integer ≧ 1, default: 4)
• context - size of the prompt context (0...4096, default: 512)

Embeddings Node

Embeddings calculation supports the following properties:

• payload - this is the actual input text
• seed - seed value for the internal pseudo random number generator (integer, default: -1, use random seed for <= 0)
• threads - number of threads to use during computation (integer ≧ 1, default: 4)
• context - size of the prompt context (0...4096, default: 512)

HTTP Endpoint Usage

Besides the sole function nodes for

• text completion,
• tokenization and
• embeddings calculation

this repository also contains example flows which answer incoming HTTP requests. The prompt itself and any desired parameters have to be passed as query parameters, the result of the called function will then be returned in the body of the HTTP response.

For inferencing, the following parameters are supported (most of them will be copied into a msg property of the same name):

• prompt - will be copied into msg.payload
• seed - will be copied into msg.seed
• threads - will be copied into msg.threads
• context - will be copied into msg.context
• keep - will be copied into msg.keep
• predict - will be copied into msg.predict
• topk - will be copied into msg.topk
• topp - will be copied into msg.topp
• temperature - will be copied into msg.temperature
• batches - will be copied into msg.batches
• grammar - will be copied into msg.grammar

Tokenization and embeddings calculation endpoints support a subset of these parameters - as required by their respective function nodes.

Examples

If you have cURL installed (if not - but you want it - just follow the instructions found in their docs) (and assuming that your Node-RED installation is listening at port 1880) you may use the following commands to "smoke test" the imported flows:

Text Completion

curl "http://127.0.0.1:1880/stablelm?prompt=who%20was%20Joseph%20Weizenbaum%3F"

Tokenization

curl "http://127.0.0.1:1880/stablelm-tokenization?prompt=who%20was%20Joseph%20Weizenbaum%3F"

Embeddings Calculation

curl "http://127.0.0.1:1880/stablelm-embeddings?prompt=who%20was%20Joseph%20Weizenbaum%3F"

License

MIT License