7 dépôts
Techniques to reduce the number of active parameters during inference to increase token throughput.
Distinct from Model Performance Optimization: Distinct from general performance optimization by focusing specifically on reducing the active parameter set (sparsity) during execution.
Explore 7 awesome GitHub repositories matching artificial intelligence & ml · Model Sparsification. Refine with filters or upvote what's useful.
pykan is a library for implementing Kolmogorov-Arnold Networks, replacing fixed node activation functions with learnable spline functions located on the network edges. It serves as an interpretable AI framework and symbolic regression tool designed to derive transparent mathematical rules from complex data. The project focuses on converting learned numerical functions into human-readable symbolic expressions through library matching and formula conversion. It utilizes additive-compositional topologies and learnable piecewise polynomial segments to approximate non-linear mappings. The framewo
Uses regularization-driven sparsification to force unimportant connections to zero for better interpretability.
PowerInfer is an inference engine and serving framework designed to run large language models on local hardware. It combines a hybrid CPU-GPU offloader, a quantization tool, and a sparse model optimizer to enable the execution of high-parameter models on consumer-grade devices. The system distinguishes itself through neuron-activation-based offloading, using a predictor model to preload frequent neurons into VRAM while keeping rare neurons in system memory. This hybrid execution model balances workloads between the GPU and CPU based on input patterns to optimize memory access and increase tok
Reduces active parameters during execution to increase token throughput and improve processing speed.
MergeKit is a toolkit for combining multiple pre-trained large language models into a single entity using algorithmic blending. It provides a specialized system for parameter interpolation and weight extraction to unify model capabilities. The project distinguishes itself through an evolutionary merge optimizer that tunes parameters based on quantitative evaluation metrics. It also features a mixture of experts orchestrator capable of converting dense models into sparse architectures and a tokenizer alignment tool for transplanting embeddings between different models. The toolkit covers a br
Implements parameter pruning and sign conflict resolution to create sparse model representations.
x-transformers est une bibliothèque PyTorch et un toolkit de recherche pour construire des architectures transformer. Il fournit un framework modulaire pour implémenter la recherche transformer expérimentale, incluant une suite de mécanismes d'attention avancés, des outils de modélisation de séquences longues et un framework pour les vision transformers. Le projet se distingue par son accent sur les composants haute performance et économes en mémoire, tels que Flash Attention avec des noyaux tuilés et l'attention multi-requêtes. Il implémente également des méthodes spécialisées pour étendre les fenêtres de contexte, incluant la récurrence de séquence et les plongements positionnels rotatifs. La bibliothèque couvre un large éventail de capacités architecturales, incluant divers schémas de normalisation pour stabiliser l'entraînement, des réseaux feedforward à portes et des topologies de couches personnalisées comme les réseaux Macaron. Elle prend en charge les constructions d'encodeur et de décodeur, fournissant des outils pour la génération de séquence autorégressive et les tâches vision-langage comme la légende d'image.
Implements top-k selection to zero out low-importance attention scores, reducing computational overhead.
Ce projet est une bibliothèque PyTorch pour construire et entraîner des réseaux Kolmogorov-Arnold. Il implémente une architecture de réseau neuronal qui remplace les fonctions d'activation fixes par des fonctions basées sur des splines apprenables sur les arêtes, servant d'outil pour le machine learning interprétable. L'implémentation utilise des opérations matricielles reformulées pour réduire la surcharge mémoire et augmenter la vitesse de calcul. Elle emploie une régularisation L1 pour sparsifier les poids du réseau, ce qui améliore la transparence de la logique interne et des décisions du modèle. Le framework couvre un éventail de capacités, incluant l'approximation de fonctions basée sur une grille, les fonctions d'activation B-spline et l'optimisation de modèles d'apprentissage profond. Ces fonctionnalités sont construites en utilisant des tenseurs PyTorch natifs pour prendre en charge la différenciation automatique et l'accélération matérielle.
Includes utilities for model weight sparsification via L1 regularization to improve interpretability.
Model-Optimizer is a deep learning toolkit and framework dedicated to compressing, pruning, quantizing, and optimizing neural network architectures. It provides methodologies covering weight quantization, model distillation, and speculative decoding for efficient text generation, alongside automated neural architecture search for discovering optimal network structures. The library implements post-training quantization pipelines that convert high-precision neural network weights into lower-bit formats using calibration data. Additional optimization techniques include teacher-student knowledge
Transforms pre-trained dense neural network models into sparse variants using magnitude-based thresholding or data-driven calibration without retraining.
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
Implements a dynamic pruning schedule to reduce the number of non-zero weights and decrease inference cost.