5 dépôts
Extends tensor parallelism across machines by joining local CUDA ranks into a global NCCL communicator coordinated by a head node.
Distinct from Tensor Parallelism: Distinct from general Tensor Parallelism: specifically addresses multi-node coordination with NCCL communicators.
Explore 5 awesome GitHub repositories matching artificial intelligence & ml · Multi-Node. Refine with filters or upvote what's useful.
mistral.rs is an inference engine for large language models that runs locally and exposes models behind OpenAI and Anthropic-compatible APIs. It serves as a multi-model serving platform, capable of loading several models in a single server process with per-request routing and on-demand loading and unloading. The engine supports multimodal inference, processing text alongside images, video, audio, and speech inputs, and includes a quantized model deployment runtime that reduces memory use and speeds up inference on consumer hardware. The project distinguishes itself through an agentic tool exe
Extends tensor parallelism across machines by joining local CUDA ranks into a global NCCL communicator.
Dynamo is a distributed inference orchestration platform designed for large language models. It functions as a system to coordinate prefill and decode phases across GPU nodes, utilizing a multi-backend runtime adapter to connect engines like vLLM and TensorRT-LLM through a unified block-oriented memory interface. An OpenAI-compatible API server provides the frontend for integration with existing tools and clients. The project is distinguished by its disaggregated serving architecture, which separates prompt processing and token generation onto independent GPU pools to optimize throughput and
Spreads tensor-parallel inference across multiple hardware nodes using global NCCL communicators.
FlashInfer is a library of high-performance GPU kernels purpose-built for accelerating large language model inference. It provides optimized implementations for attention operations (including flash attention, page attention, multi-head latent attention, and cascade attention) using paged key-value caches, fused kernel composition, and just-in-time compilation. The library also includes specialized kernels for mixture-of-experts layers, block-scaled low-precision quantization (FP8, FP4), and distributed collective communication. What distinguishes FlashInfer is its fused all-reduce communicat
Provides GPU buffer allocation with multicast access for multi-node communication in distributed inference.
NCCL est une bibliothèque de communication haute performance et un framework de calcul GPU distribué conçu pour exécuter des échanges de données collectifs et point à point sur plusieurs GPU dans des systèmes à un ou plusieurs nœuds. Il sert de couche de transport GPU RDMA et d'orchestrateur de mémoire, facilitant la synchronisation à large bande passante des données et des gradients de modèle pour l'entraînement et l'inférence GPU distribués. La bibliothèque se distingue par sa capacité à exécuter des primitives de communication directement depuis les noyaux (kernels) GPU, supprimant le CPU hôte du chemin critique. Elle utilise une sélection de chemin consciente de la topologie pour optimiser le mouvement des données et emploie un transport réseau basé sur RDMA, incluant InfiniBand et NVLink, pour permettre un accès mémoire zéro-copie entre les appareils sur différents nœuds physiques. Le projet couvre un large éventail de modèles de communication collective, notamment les réductions, les diffusions (broadcasts), les rassemblements (gathers) et les échanges tous-à-tous, ainsi que l'accès mémoire distant point à point. Il fournit une gestion complète des communicateurs pour initialiser, partitionner et redimensionner les groupes GPU, ainsi qu'une gestion spécialisée de la mémoire pour enregistrer les tampons (buffers) et coordonner la mémoire partagée des appareils. Le système inclut une suite d'outils de surveillance et d'observabilité pour le suivi de la santé, la journalisation diagnostique et la surveillance des événements en temps réel, ainsi que des interfaces d'intégration pour les frameworks de machine learning, les graphes CUDA, MPI et Python.
Coordinates GPU communication across multiple physical nodes by establishing NVLink domains and identifiers.
Distributed-llama is a distributed inference engine and command line tool for running large language models across multiple networked machines. It functions as a compute cluster manager that coordinates worker nodes to share the computational load of a single model. The system utilizes tensor parallelism to shard model weights across different hosts, allowing the execution of models that exceed the memory capacity of a single piece of hardware. It includes a dedicated format converter to transform standard model files into a compatible binary layout optimized for distributed loading. The eng
Groups independent machines into a virtual compute resource to enable multi-node tensor parallelism.