Slide decks

What Dynamo is, who runs it, the open-source community, and a tour of every component with published results.

Hands-on setup from a laptop to Kubernetes, tested recipes, the tuning levers that matter, and agent skills that optimize for you.

Where Dynamo fits in RL post-training today, what customers are doing with it, and the roadmap across three pillars.

How Dynamo serves agent programs: session IDs, tracing and simulation, program-level scheduling with ThunderAgent, and programmatic KV caching.

Vision-language, speech and diffusion models on one OpenAI-compatible serving stack, with live demos.

Faster model loading and worker startup with ModelExpress and snapshots.

Distributed KV cache management: the Dynamo router plans where KV goes, and KVCR moves it between workers and keeps it through failures.

Simulating the whole inference stack (traffic, scheduler, KV cache, router, planner, GPU timing) on a laptop, without GPUs.

Take Dynamo out of the engine process: a CPU-only sidecar plugs a stock vLLM, SGLang or TRT-LLM server into Dynamo.

How the Dynamo router picks a worker, survives overload and restarts, decides what gets admitted, and how to plug in your own policy — plus where it's going.

The existing router run one level up, where each pool acts as a single super-worker.

Keep serving when something breaks.

Planner decides how many engines to run; Grove decides what one engine instance is and where it lands.

Benchmark any inference endpoint with production-shaped traffic: architecture, load control, SLA search, goodput, outputs, and who uses it.

One API that finds the fastest validated backend for every module of a PyTorch model: architecture, strategies, JIT vs AOT and performance.

Run models bigger than your GPU: weights stream from host RAM just in time, when it's free, how it's built, and results on video and LLM workloads.