3 repositorios
Automatically determining the optimal percentage of parameters to remove to meet resource budgets.
Distinct from Model Pruning: Focuses on the search for optimal pruning ratios rather than the execution of the pruning.
Explore 3 awesome GitHub repositories matching artificial intelligence & ml · Pruning Ratio Optimization. Refine with filters or upvote what's useful.
Torch-Pruning is a deep learning model pruning tool and neural network optimization toolkit designed for PyTorch. The library analyzes neural network graphs to remove redundant parameters and channels, reducing model size and computational costs. The framework traces computational graphs dynamically using sample inputs to map relational dependencies and identify coupled layers that require simultaneous pruning. It evaluates and ranks layer parameters across the entire network using global importance ranking and applies structural constraints to optimize overall outcomes. Additional capabili
Perform global importance ranking across network layers and apply structural constraints to optimize overall pruning results.
Este proyecto es un framework para alinear modelos de lenguaje extensos (LLM) con las preferencias humanas. Proporciona una librería para optimizar el comportamiento del modelo mapeando los datos de preferencia directamente a un objetivo de política, evitando la necesidad de un modelo de recompensa independiente. El framework utiliza un objetivo matemático de forma cerrada para ajustar los pesos del modelo maximizando la probabilidad de las respuestas preferidas frente a las no preferidas. Incluye un kit de herramientas para el ajuste fino supervisado (SFT), lo que permite a los usuarios establecer una línea base estable con datos de demostración antes de aplicar técnicas de optimización basadas en preferencias. La librería admite entrenamiento distribuido, lo que permite escalar la alineación del modelo en múltiples aceleradores de hardware mediante estrategias de fragmentación de datos y paralelización. También proporciona utilidades para preparar datasets de preferencia personalizados, mapeando pares de prompts y respuestas para estructurar la retroalimentación humana para el pipeline de entrenamiento.
Adjusts model weights by maximizing the likelihood of preferred responses relative to dispreferred ones using a closed-form mathematical objective.
PocketFlow is an integrated toolkit for deep learning model compression, distributed training, and mobile format optimization. It provides a system for reducing the size and complexity of neural networks to improve inference efficiency, featuring a dedicated engine for knowledge distillation and a mobile model optimizer. The framework differentiates itself through an automated hyperparameter tuning system that uses reinforcement learning and statistical models to determine optimal compression ratios and layer-wise bit allocation. It also includes a distributed training system that utilizes mu
Uses reinforcement learning to find optimal compression ratios that meet specific computation or FLOPs budgets.