What is introduced here is a pre-trained model with powerful capabilities, characterized by its open-access nature and specialized design for code optimization tasks.
The LLM compiler in this set has a unique construction foundation, built upon Code Llama. Through this construction, it further enhances its understanding of compiler intermediate representations (IRs), assembly languages, and optimization techniques. During the model's training process, it utilized an extremely large corpus. This corpus contains up to 546 billion LLVM-IR and assembly code annotations, offering abundant learning materials for the model.
Moreover, the model has undergone the important process of instruction fine-tuning. Through instruction fine-tuning, it can better interpret the behavior of compilers. This enables the model to more accurately understand various operations and intentions of compilers when dealing with code optimization tasks, thereby providing more efficient and precise support for code optimization.