Compiler Model Specialization for Reproducible ML Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning (ML) based compiler optimization faces issues with non-reproducible compilation results and performance variation due to lack of training data and facilities at the user side, making it difficult to achieve better performance.
Innovation Solution
A model specialization method for ML based compiler optimization that generates a specialized model using reinforcement learning (RL) or supervised learning based compiler optimization, allowing for reproducible and stable performance by utilizing an original model derived from the developer side.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the original model is directly sent from developer side to user side for inference, then the compilation process is simple, but the compilation results are not reproducible and performance varies dramatically
Solution Approach 1:
The patent applies preliminary action by pre-processing the original model into an intermediate representation format that preserves the essential characteristics needed for reproduction. The developer side performs model specialization in advance, creating a standardized intermediate form that can be reliably reconstructed at the user side without requiring access to original training data or facilities.
Solution Approach 2:
The patent introduces an intermediate representation as a mediator between the developer side and user side. This intermediate form acts as a bridge that carries the necessary model information in a standardized, reproducible format, enabling consistent compilation results across different environments without direct transfer of the original trained model.
2Productivity
If retraining the original model at the user side is performed to achieve better performance, then performance improvement may be achieved, but it requires original training data and training facilities which are not available
Solution Approach 1:
The patent extracts the essential performance-optimizing characteristics from the original trained model and embeds them into the intermediate representation. This extraction process captures the key optimization patterns and decision rules without requiring the full training dataset or training infrastructure, enabling performance improvement while avoiding the need for complete model retraining.
Solution Approach 2:
The patent transforms the model into a different parameter representation through intermediate form conversion. By changing the representation parameters from trained weights to optimization rules and patterns, the system enables performance optimization at the user side using only the intermediate representation, without needing to access original training parameters or facilities.
3Reliability
If model specialization is performed to achieve reproducible results, then compilation result reproducibility improves, but additional processing steps are required
Solution Approach 1:
The patent performs model specialization as a preliminary action at the developer side, converting the original model into an intermediate representation that inherently contains all necessary information for reproducible compilation. This advance processing eliminates the need for complex specialisation steps at the user side, as the reproducibility is built into the intermediate form itself.
Data Source
AI summary
A model specialization method for machine learning (ML) based compiler optimization includes: compiling a program by a compiler, wherein the compiler is executed by a processor; generating a specialized model according to an original model derived from a developer side and one of reinforcement learning (RL) based compiler optimization and supervised learning based compiler optimization; and utilizing the specialized model for compilation, and returning executables.