Sakana AI’s New Agent Framework Can Improve Model Deployment Speed

The Tokyo-based Artificial Intelligence (AI) firm, Sakana AI introduced a new Artificial Intelligence (AI) agent framework, which can improve the growth and deployment speed of large language models (LLM). Announced on Thursday, the company unveiled the AI ​​Gut Engineer that improves both pre-training and estimate speed of the AI ​​model by adapting the codebase. The AI ​​firm highlighted that the entire process is operated by AI agents and end-to-end. In particular, Sakan AI introduced the AI ​​scientist last year that scientific research could do.

Sakan Ai unveiled Ai Gud Engineer

One in PostThe Japanese AI firm said that after developing the AI ​​system, who can create new models, and AI can completely automate the research process, it began working on ways to intensify the deployment and estimate speed of LLM. .

The company said that research led the development of AI CUDA engineer. This is a fully automated, broad agent framework for CUDA (Compute Unified device architecture) kernel discovery and optimization.

CUDA kernels can be understood as special functions running on Nvidia GPU, allowing parallel execution of code in many threads. Due to equality, it is more customized than traditional methods and allows for acceleration of computational functions, especially with large datasets. For example, it is considered a great way to optimize the deployment and estimate of the AI ​​model.

Sakana AI said that AI Cuda engineer can automatically convert the pytorch module into Cuda Kernels, so that the deployment speedup can be greatly improved. This can produce the kernels that are called 10–100 times faster than their pitoch counterpart.

The process includes four stages. First, the agent framework converts the pytorch code into a kernel. Then, the agent only applies adaptation techniques to ensure the best kernels. Then, kernel crossover signals are added, which add several customized kernels to create new kernels. Finally, the AI ​​agent protects the high-demonstration Cuda kernels in a collection, which is used to improve performance. The company has also published Study This gives details of further process.

Along with the paper, Sakan AI AI AI Cuda is also publishing Engineer Archive, a dataset with more than 30,000 kernels generated by AI. These kernels are issued under the CC-BY-4.0 license and can be reached through the face of the hugging face.

Additionally, the Japanese firm also launched a website, which allows visitors to explore 17,000 verified kernels and their profiles. The website allows users to detect these kernels in 230 tasks, and also allows them to compare Cuda kernels in personal experiments.