ai-dynamo/dynamo

Features

• Disaggregated Inference Orchestration - Coordinates separate prefill and decode worker pools and manages the necessary KV cache transfers between them.
• Prefill-Decode Disaggregation - Separates prompt processing and token generation onto independent GPU pools to optimize throughput and memory.
• OpenAI-Compatible APIs - Exposes a standard HTTP frontend with completions endpoints for compatibility with existing LLM tools.
• Activation and KV Cache Offloaders - Enables remote access and reuse of key-value cache blocks across nodes via high-speed interconnects.
• KV-Cache-Aware Request Routing - Routes inference requests to GPUs based on KV cache affinity to avoid redundant computation.
• LLM Serving Architectures - Implements high-performance architectures to coordinate API gateways and worker nodes for large-scale LLM serving.
• Inference Orchestration - Provides a system for scaling and managing the distribution of prefill and decode inference workloads across multiple GPU nodes.
• Disaggregated Phase Scaling - Adjusts the number of workers dedicated to prefill and decode phases separately based on real-time metrics.
• Disaggregated Inference - Implements an architecture that separates prefill and decode stages across distinct hardware nodes to optimize throughput.
• Multi-Backend Inference Orchestration - Routes requests across various backend engines to manage serving and resource allocation within a single orchestration layer.
• Prefix-Aware Routing - Directs requests to workers holding required prompt prefixes in memory to accelerate response times.
• Token-State Continuation - Tracks accumulated token sequences mid-generation so a new worker can resume from the exact failure point.
• Key-Value Cache Reuse - Retains and reuses key-value cache entries across multiple requests to accelerate repeated prompts.
• Worker Cache State Tracking - Consumes cache events from backend workers to maintain an accurate view of cached blocks for optimized routing.
• KV Cache Management - Allocates memory blocks across storage tiers to optimize the key-value cache and reduce prompt recomputation.
• Remote Cache Block Sharing - Exchanges block metadata and access permissions across nodes via high-speed interconnects for remote cache reuse.
• Inference Runtime Integrations - Integrates various inference engines into a distributed runtime for language, embedding, and multimodal models.
• Autoscaling Systems - Profiles workloads and right-sizes GPU pools to meet specific latency SLAs while minimizing cost.
• RDMA GPU Transfers - Moves KV cache blocks directly between GPU VRAMs over RDMA or NVLink without blocking forward passes.
• KV Cache VRAM Transfers - Moves KV cache from prefill VRAM to decode VRAM without blocking ongoing GPU passes.
• Inference - Implements a translation layer that connects vLLM and TensorRT-LLM into a unified block-oriented memory interface.
• Inference Engine Adapters - Implements a multi-backend runtime adapter to connect different inference engines through a unified block-oriented memory interface.
• Paged KV Cache Management - Implements a paged key-value cache that pools and reuses fixed-size blocks across requests.
• Master-Worker Coordination - Coordinates the lifecycle and task assignment of inference engine workers across multiple backends.
• Fault Tolerance - Provides systems for managing failures and ensuring resilience in distributed inference applications.
• Resource Deployment Planning - Allocates compute resources across prefill and decode nodes to meet specified latency targets.
• Workload Simulations - Mimics backend API behavior and synthetic traffic patterns to validate routing and infrastructure logic without consuming GPUs.
• External Tool Execution - Coordinates communication between language models and external functions to execute programmatic tasks.
• Inference Pipeline Observability - Exposes metrics, distributed traces, and dashboards for the complete inference pipeline.
• Diffusion Models - Executes inference for language, embedding, vision, and diffusion models to generate text, images, or video.
• Hardware Acceleration Support - Optimizes inference workloads for the latest NVIDIA GB200 hardware architectures.
• Agentic Latency Optimizations - Implements per-request priorities and cache pinning to optimize performance for autonomous agent-driven workloads.
• Warm Pool and Predictive Scaling - Forecasts request counts and sequence lengths using time-series methods to proactively scale GPU resources.
• Multimodal Inference Engines - Processes images, video, and audio alongside text across multiple backend engines for complex workloads.
• Speculative Decoding Strategies - Uses a smaller draft model to propose candidate tokens that a larger target model verifies in parallel.
• Worker Availability Verification - Tracks active and ready inference nodes to ensure the router only directs traffic to healthy workers.
• Adapter-Aware Routing - Routes requests to specific GPU nodes that have the required LoRA adaptation weights already loaded in memory.
• Dynamic Adapter Swapping - Swaps LoRA adapters in and out at runtime from compatible storage without restarting the engine.
• Remote Weight Streaming - Transfers model weights directly between GPUs over high-speed interconnects to reduce cold-start latency.
• LoRA Adapter Interfaces - Provides interfaces for dynamically loading and switching between different LoRA adapters during inference.
• Multi-Node - Spreads tensor-parallel inference across multiple hardware nodes using global NCCL communicators.
• Token State Continuations - Captures and transfers token state from failed workers to resume generation from the exact failure point.
• Disaggregated Serving - Separates prompt processing and token generation into distinct worker pools to optimize hardware utilization.
• State Snapshots - Captures system state snapshots to enable recovery and consistent state replication across the cluster.
• AI Memory Tiering - Manages block reuse and eviction across HBM, DRAM, and NVMe tiers to balance speed and capacity.
• Load-Aware Request Routing - Routes inference requests to the least-loaded worker based on active cache utilization.
• Auto-scaling Engines - Dynamically adjusts replica counts for prefill and decode engines based on traffic signals.
• Inference Worker Recovery - Detects unhealthy workers, drains in-flight requests, and reroutes traffic to maintain service continuity.
• Capacity Scaling - Computes scaling targets for GPU workers based on real-time throughput and load metrics.
• Inference Scaling Analysis - Executes concurrency sweeps to identify saturation points and generate performance curves for the inference stack.
• Topology-Aware Schedulers - Employs a topology-aware operator to optimally place interdependent inference components across racks and hosts.
• Hardware Profile Deployments - Automatically profiles and deploys the optimal engine configuration based on specified model, hardware, and target requirements.
• SLA-Driven Resource Planning - Simulates thousands of serving configurations offline to select optimal resource allocations for given SLAs.
• Deployment Simulators - Evaluates thousands of serving configurations offline to find the optimal setup without consuming GPU resources.
• Scaling Profiles - Generates performance data and selects optimal engine configurations for the system planner.
• Cache-Utilization Scaling - Uses static thresholds on queue depth and cache utilization to scale engines without strict SLA targets.
• Kubernetes Orchestrators - Orchestrates GPU-accelerated inference workloads in Kubernetes clusters using topology-aware scheduling.
• Kubernetes Deployments - Launches and manages distributed inference workloads on Kubernetes clusters.
• LLM Deployment Operators - Ships a controller that manages inference workloads using custom resource definitions and topology-aware scheduling on Kubernetes.
• Kubernetes Operators - Provides a Kubernetes operator to manage the deployment and scaling of inference workloads using custom resource definitions.
• Adapter Management - Dynamically loads and removes fine-tuned LoRA adapters from storage without restarting the inference engine.
• Queue Reordering - Controls request reordering in the router queue using policies like first-come or shortest-processing-time.
• Inference Load Balancers - Distributes incoming requests among backend workers using configurable scheduling policies.
• Data-Parallel Rank Balancing - Distributes requests across data-parallel ranks using external control for hybrid load balancing.
• Dynamic Capacity Adjustment - Allows runtime modification of worker pool capacity for prefill and decode phases without requiring application restarts.
• Independent Model Component Scaling - Runs the vision encoder as a separate worker that scales independently from the language model.
• Token Streaming - Provides real-time delivery of generated tokens to the client to minimize perceived latency.
• Multi-Architecture Support - Provides execution support across a wide range of NVIDIA GPU architectures from Ampere through Blackwell.
• Inference Request Queuing - Manages the order and dispatch of concurrent inference requests to optimize time-to-first-token.
• Generation State Migration - Tracks accumulated token sequences mid-generation to allow a new worker to resume from the exact failure point.
• Hierarchical Request Routing - Routes requests through a control layer to separate pools optimized for different request classes.
• LLM Performance Analyzers - Ships tools for identifying execution and logic bottlenecks specifically within language model serving configurations.
• Request Migration Monitors - Tracks request migrations and sequence length violations to identify reliability patterns in the distributed serving layer.
• Inference Performance Monitoring - Tracks model-serving metrics such as GPU utilization, request throughput, and response latency.
• Inference Metric Recording - Exposes engine-level metrics and monitoring data, including token counts and confidence scores, for the serving pipeline.
• Performance Visualization - Provides visual dashboards and plots to compare model and hardware configurations using telemetry data.
• Kubernetes Custom Resource Definitions - Uses Kubernetes Custom Resource Definitions to declaratively manage worker groups and topology placement.
• Inference Latency Targets - Uses estimates and tuning to scale engines toward precise time-to-first-token and inter-token latency targets.
• Performance Profiling - Measures engine performance data before deployment to determine hardware allocation and scaling requirements.
• Request Migrations - Moves active requests to healthy workers during failures to prevent request loss.
• Data-Parallel Rank Routing - Directs requests to a specific data-parallel replica based on rank assignment to ensure consistent reuse.
• Inference Engines - Distributed inference serving framework for datacenter scales.
• Model Serving & Deployment - Optimizes inference for generative AI in distributed environments.

Istoric stele