Rock solid AI – Granite on a mainframe

Let’s start with what people are familiar with, ChatGPT. ChatGPT is a highly-trained and clever chatbot. The GPT part of its name stands for Generative Pre-trained Transformer. Generative means that it can generate text or other forms of output. Pre-trained means that it has been trained on a large dataset. And Transformer refers to a type of neural network architecture enabling it to understand the relationships and dependencies between words in a piece of text. IBM’s Granite 3.0 is very similar to ChatGPT, except that it is optimized for specific enterprise applications rather than general queries.

Just a side note, I was wondering about the choice of name for the product. In the UK, the traditional gift for a 90^th anniversary is granite. I just wondered whether there was some kind of link. In 1933 IBM bought Electromatic Typewriters, but I can’t see the link. Or maybe I’ve been doing too many brain-training quizzes!

Granite was originally developed by IBM and intended for use on Watsonx along with other models. In May this year, IBM released the source code of four variations of Granite Code Models under Apache 2, allowing completely free use, modification, and sharing of the software.

In the original press release in September 2023, IBM said: “Recognizing that a single model will not fit the unique needs of every business use case, the Granite models are being developed in different sizes. These IBM models – built on a decoder-only architecture – aim to help businesses scale AI. For instance, businesses can use them to apply retrieval augmented generation for searching enterprise knowledge bases to generate tailored responses to customer inquiries; use summarization to condense long-form content – like contracts or call transcripts – into short descriptions; and deploy insight extraction and classification to determine factors like customer sentiment.”

The two sizes mentioned in that press release are the 8B and 2B models.

In October this year, Version 3.0 was released, which is made up of a number of models. In fact the press release tells us that “IBM Granite 3.0 release comprises: 

• Dense, general purpose LLMs: Granite-3.0-8B-Instruct, Granite-3.0-8B-Base, Granite-3.0-2B-Instruct and Granite-3.0-2B-Base.
• LLM-based input-output guardrail models: Granite-Guardian-3.0-8B, Granite-Guardian-3.0-2B
• Mixture of experts (MoE) models for minimum latency: Granite-3.0-3B-A800M-Instruct, Granite-3.0-1B-A400M-Instruct
• Speculative decoder for increased inference speed and efficiency: Granite-3.0-8B-Instruct-Accelerator.

Let’s put a little more flesh on the bones of those models:

• The base and instruction-tuned language models are designed for agentic workflows, Retrieval Augmented Generation (RAG), text summarization, text analytics and extraction, classification, and content generation.
• The decoder-only models are designed for code generative tasks, including code generation, code explanation, and code editing, and are trained with code written in 116 programming languages.
• The time series models are lightweight and pre-trained for time-series forecasting, and are optimized to run efficiently across a range of hardware configurations.
• Granite Guardian can safeguard AI by ensuring enterprise data security and mitigating risks across a variety of user prompts and LLM responses.
• Granite for geospatial data is an AI Foundation Model for Earth Observations created by NASA and IBM. It uses large-scale satellite and remote sensing data.

In case you didn’t know, agentic workflows refer to autonomous AI agents dynamically interacting with large language models (LLMs) to complete complex tasks and produce outputs that are orchestrated as part of a larger end-to-end business process automation.

Users can deploy open-source Granite models in production with Red Hat Enterprise Linux AI and watsonx, at scale. Users can build faster with capabilities such as tool-calling, 12 languages, multi-modal adaptors (coming soon), and more, IBM tells us.

IBM is claiming that Granite 3.0 is cheaper to use compared to previous versions and other LLM (large language models) such as GPT-4 and Llama

IBM also tested the Granite Guardian against other guardrail models in terms of their ability to detect and avoid harmful information, violence, explicit content, substance abuse, and personal identifying information, showing it made AI applications safer and more trusted.

We’re told that the Granite code models range from 3 billion to 34 billion parameters and have been trained on 116 programming languages and 3 to 4 terabytes of tokens, combining extensive code data and natural language datasets. If you want to get your hands on them, the models are available from Hugging Face, GitHub, Watsonx.ai, and Red Hat Enterprise Linux (RHEL) AI. A curated set of the Granite 3.0 models can be found on Ollama and Replicate.

At the same time, IBM released a new version of watsonx Code Assistant for application development. The product leverages Granite models to augment developer skill sets, simplifying and automating their development and modernization efforts. It simplifies and accelerates coding workflows across Python, Java, C, C++, Go, JavaScript, Typescript and more.

Users can download the IBM Granite.Code (which is part of the watsonx Code Assistant product portfolio) extension for Visual Studio Code to unlock the full potential of the Granite code model from here.

It seems to me that the Granite product line is a great way for organizations to make use of AI both on and off the mainframe. I’m looking forward to seeing what they announce with Granite 4.0 and other future versions.