DeciCoder 6B
About
DeciCoder 6B is a state-of-the-art large language model tailored for code generation, featuring 6 billion parameters and a transformer decoder architecture. It utilizes variable Grouped Query Attention and has a 2,000-token context window, enhancing its ability to generate and complete code in languages such as Python, Java, JavaScript, Rust, C++, C#, and Go. Particularly adept at code completion and generation, it demonstrates superior performance on the MultiPL HumanEval benchmark, surpassing many larger models. DeciCoder 6B is designed for cost efficiency and scalability, with a throughput 19 times higher than some competitors on Qualcomm's Cloud AI 100 platform. The model's architecture benefits from Deci's proprietary Neural Architecture Search technology (AutoNAC), enabling dynamic attention pattern adjustments for optimal performance and computational efficiency 127.