allm 1.0.7

A simple and efficient python library for fast inference of LLMs

ALLM

ALLM is a Python library designed for fast inference of GGUF (Generic Global Unsupervised Features) Large Language Models (LLMs) on both CPU and GPU. It provides a convenient interface for loading pre-trained GGUF models and performing inference using them. This library is ideal for applications where quick response times are crucial, such as chatbots, text generation, and more.

Features

• Efficient Inference: ALLM leverages the power of LLM models to provide fast and accurate inference.
• CPU and GPU Support: The library is optimized for both CPU and GPU, allowing you to choose the best hardware for your application.
• Simple Interface: With a straightforward command line support, you can easily load models and perform inference with just a single command.
• Flexible Configuration: Customize inference settings such as temperature and model path to suit your needs.
• Automated Hosting Configuration: Models are swiftly downloaded and configured in your environment, enabling them to be operational within minutes.

Operating System Compatibility

This table outlines the compatibility of different operating systems with their respective providers:

OS Type	Windows	Linux	MacOS
Status	Supported	Supported	Coming Soon
Dependencies	Python, VS Tools	Python, bash sudo apt-get install build-essential -y	Coming Soon
Models Support	- Local Models - VertixAI - Azure OpenAI	- Local Models - VertixAI - Azure OpenAI	Coming Soon

|

Supported Models

LLM Family	Hosting	Supported LLMs
ALLM	Self Hosted (gguf)	Mistral, Mistral_instruct, Llama2, full list availble in supported model section.
Azure	AzureOpen AI	gpt-35-turbo, gpt-4, gpt-4-turbo or any.
Google LLMs	VertexAI deployment	gemini-pro, text-bison@001 or any.
Llama2	Azure deployment	llama2-7b, llama2-13b, llama2-70b
Mistral	Azure deployment	Mistral-7b, Mixtral-7bx8

Installation

You can install ALLM using pip:

pip install allm

Usage

0.1 LocalModel Generic Prompt

You can start inference with a simple 'allm-run' command. The command takes name or path, temperature(optional), max new tokens(optional) and additional model kwargs(optional) as arguments.

when you run the allm-run, Defualt Mistral model will be downloaded to your systems and gets configured automatically, if the model name is not provided.

allm-run --name model_name_or_path

API

0.2 LocalModel Generic API

You can initiate the inference API by simply using the 'allm-serve' command. This command launches the API server on the default host, 127.0.0.1:5000. If you prefer to run the API server on a different port and host, you have the option to customize the apiconfig.txt file within your model directory.

allm-serve

ALLM Agents

1.1 New Agent Creation

To create local agent, begin by loading your knowledge documents into the database using the allm-newagent command and specifying the agent name:

allm-newagent --doc "document_path" --agent agent_name

or

allm-newagent --dir "directory containing files to be ingested" --agent agent_name

1.2 Agent Chat

After agent is created successfully with your knowledge document, you can start the local agent chat with the allm-agentchat command:

allm-agentchat --agent agent name

After your agents are created you can also initiate agent-specific API server using the allm-agentapi command:

1.3 Agent API

allm-agentapi --agent agent name

After your agents are created you can also update the knowledge on the existing agent by adding documents using allm-updateagent command:

allm-updateagent --agent agent name

Supported Cloud models.

ALLM supports all types of Generative LLMs on AzureOpenAI & VertexAI, including GPT(s) & Geminipro models. You can start local inference of cloud based models using the following command:

2.1 VertixAI Generic Prompt

allm-run-vertex --projectid Id_of_your_GCP_project --region location_of_your_cloud_server

or

allm-run-vertex

2.2 AzureOpneAI Generic Prompt

allm-run-azure --key key --version version --endpoint https://{your_endpoint}.openai.azure.com --model model_name

or

allm-run-azure

You can also have a custom agent working with your cloud deployed model using the following command. It is important to note that before this step, agent should be created using section: 1.1 New Agent Creation.

2.3 VertixAI AgentChat

allm-agentchat-vertex --projectid Id_of_your_GCP_project --region location_of_your_cloud_server --agent agent_name

model\vertex-config.json needs to be configured to use below command, this will ensure projectid, region are captured.

allm-agentchat-vertex --agent agent_name

2.4 AzureOpenAI AgentChat

allm-agentchat-azure --key key --version version --endpoint https://{your_endpoint}.openai.azure.com --model model_name --agent agentname

model\azure-config.json needs to be configured to use below command, this will ensure endpoint, modelname etc are captured.

allm-agentchat-azure --agent agentname

model_name is an optional parameter in both vertex and azure, if not mentioned, inference will work on gemini-1.0-pro-002 for vertex and gpt-35-turbo for OpenAI by default.

2.5 AzureOpenAI AgentChat API

allm-agentapi-azure --agent agentname

2.6 AzureOpenAI AgentChat API

allm-agentapi-vertex --agent agent_name

2.7 ALLM-Enterprise

You can launch the UI with the following command:

allm-launch

Supported Model names.

• Llama3
• Llama2
• Llama
• Llama2_chat
• Llama_chat
• Mistral
• Mistral_instruct