Meta Platforms Inc.’s artificial intelligence research team said today it’s open-sourcing a suite of robust AI models called the Meta Large Language Model Compiler.

According to the researchers, it can transform the way developers go about code optimization for LLM development, making the process faster and more cost-efficient.

In a blog post, Meta’s systems research team explains that training LLMs is a resource-intensive and hugely expensive task that involves extensive data collection and large numbers of graphics processing units. As such, the process is prohibitive for many organizations and researchers.

However, the team believes that LLMs can help to simplify the LLM training process through their application in code and compiler optimization, which refers to the process of modifying software systems to make them work more efficiently or use fewer resources.

The researchers said the idea of using LLMs for code and compiler optimization is one that has been underexplored. So they set about training the LLM Compiler on a massive corpus of 546 billion tokens of LLVM Project and assembly code, with the purpose of making it able to “comprehend compiler intermediate representations, assembly language and optimization techniques.”

In the paper, Meta’s researchers wrote that the LLM Compiler’s enhanced comprehension of those techniques enables it to perform tasks that could previously only be done by humans or specialized tools.

Furthermore, they claim that the LLM Compiler has demonstrated enormous efficiency in code size optimization, achieving 77% of the optimizing potential of an autotuning search in their experiments. They say this shows its potential to substantially reduce code compilation times and enhance code efficiency across various applications.

The LLM Compiler achieved even better results when challenged with code disassembly tasks. It scored a 45% success rate in round-trip disassembly, with 14% exact matches, when challenged to convert x86_64 and ARM assembly back into LLMV-IR, demonstrating its potential for tasks such as legacy code maintenance and reverse engineering of software.

“LLM Compiler paves the way for exploring the untapped potential of LLMs in the realm of code and compiler optimization,” said Cris Cummins, one of the core contributors to the project.

Meta’s team believes that LLM Model Compiler can potentially enhance many aspects of software development. For instance, researchers gain more avenues for exploring AI-powered compiler optimizations, while software developers could realize faster code compilation times, create more efficient code and even build new tools for understanding and fine-tuning complex applications and systems.

To help make this happen, Meta said it’s releasing the LLM Compiler under a permissive commercial license, which means both academic researchers and companies can use it and adapt it in any way they see fit.

Although encouraging in some ways, the LLM Compiler raises questions about the evolution of software design and development and the role of human software engineers. It delivers much more than just incremental efficiency gains, representing a fundamental shift in the way code and compiler optimization technology is approached.

Image: SiliconANGLE/Microsoft Designer