Nvidia Dynamo Day Baseten Inference Stack

The Baseten Inference Stack at NVIDIA Dynamo Day Announcing our Series F . Learn more

Community

The Baseten Inference Stack at NVIDIA Dynamo Day

How NVIDIA Dynamo fits into the Baseten Inference Stack, plus highlights from what we shared at Dynamo Day.

Authors

Rachel Rapp

Last updated February 3, 2026

Share

Update: NVIDIA announced Dynamo 1.0 at GTC! The Baseten Inference Stack powers our Dedicated Inference and Model APIs, delivering the lowest latency, highest throughput, and highest uptime for generative AI workloads. We achieve this by combining open-source and in-house tools—runtime engines, scheduling layers, and model-specific optimizations—and using each as best suited to a particular model and use case. One of those tools is NVIDIA Dynamo . We like Dynamo because it’s framework-agnostic, so it doesn’t stop us from choosing the best inference engine (TensorRT-LLM/SGLang/vLLM) for a specific model and use case. We also love that it’s open-source, with biweekly releases and ~90 commits per week (our engineers are frequent contributors to the Dynamo and TensorRT-LLM ecosystems, shipping fixes , new features , and performance improvements —we also just open-sourced our Suffix Automaton MTP Accelerator from the Baseten Speculation Engine here ). We were excited to present our work using NVIDIA Dynamo at massive scale at Dynamo Day for thousands of attendees. Here are some of the highlights of the event and features we’re looking forward to working more closely with. ✕ Screenshot from Baseten’s Lead DevRel, Philip Kiely’s talk at Dynamo Day on how we use NVIDIA Dynamo as part of the Baseten Inference Stack (along with different inference engines including TensorRT-LLM, SGLang, and vLLM). NVIDIA Dynamo Day: The highlights For those needing a primer (or a refresher), NVIDIA Dynamo is an inference framework built for serving and scaling generative AI workloads in distributed environments. We use Dynamo daily for our customers in production, and our engineers take those learnings to improve Dynamo itself and contribute back to the project (some examples here and here ). As a result, we achieve metrics like a 50% reduction in time to first token (TTFT), 34% reduction in time per output token (TPOT), and 61% higher throughput on workloads with long inputs and outputs (check out the benchmarks in our blog here ). Dynamo is all about improving system-level inference performance via three main optimizations: Disaggregated serving

KV cache-aware routing

KV cache offloading

The whole is greater than the sum of its parts (pardon the cliché)—each optimization compounds the performance gains of the others when used together. At Dynamo Day, technical leads, engineers, and product managers covered topics including: The challenges of serving Mixture of Expert-architecture models at scale

How Dynamo works under the hood + key features and community growth

The OSS inference ecosystem featuring vLLM, SGLang, llm-d, and TensorRT-LLM

Learnings from Dynamo users (featuring our Lead DevRel, Philip Kiely , alongside technical team members from Pinterest and Prime Intellect)

The state of inference today and upcoming trends

NVIDIA’s team touched on several features we’re excited about—more on those below. Dynamo in The Baseten Inference Stack Features of Dynamo that we actively use (or are keen to use more heavily) include Dynamo’s native support for disaggregated prefill and decode, broader support for multimodal model serving (including recently added support for vLLM), and KV cache offloading. We’re also heavy users of Dynamo’s KV cache-aware routing— check out our blog to learn more about the 34%-62% performance gains we see while using it in production. Disaggregated serving (prefill / decode) LLM inference has two phases: prefill and decode. During the prefill phase, context is loaded and a Key-Value (KV) cache is built; during decode, a response is generated token by token. Prefill is compute-bound, and affects time to first token (TTFT)—you can think of it as the LLM “thinking” before it “speaks.” Decode is memory-bound and measured by inter-token latency (ITL) or time per output token (TPOT)—this is the LLM’s generation speed. When prefill and decode are done together on a single GPU, long, compute-heavy prefill work can block latency-sensitive token decoding and lead to poor GPU utilization under load. Disaggregating prefill and decode phases (i.e., assigning them to different GPUs) means you can optimize for both TTFT and TPOT by scaling and optimizing for each phase separately, like applying tensor parallelism for prefill and expert parallelism for decoding. By removing the bottleneck, disaggregated serving can also massively increase throughput (tokens per second, TPS)—up to 6x higher TPS per GPU is realistic. ✕ Screenshot from Harry Kim’s talk “State of NVIDIA Dynamo” explaining disaggregated serving. Dynamo includes an AIConfigurator that suggests good configurations for disaggregated serving with SGLang, TRT-LLM, and vLLM, given a target model, hardware, input and output lengths, and your TTFT and ITL/TPOT SLAs. Similar to our KV cache-aware routing post , we have new benchmarks coming out in the next few weeks on the throughput gains we see using disaggregated serving for high-traffic workloads—stay tuned! Multimodal model serving Building on disaggregated serving and text-based KV cache reuse, Dynamo provides native support for an Encode → Prefill → Decode (EPD) pattern that can encode images as well as text for multimodal models. Support for an embedding cache is also expected in the coming months, which will further reduce latency (TTFT) by skipping encoding repeated inputs (such as images/videos). As we scale up our multimodal features, we’re excited to use these in our Dedicated Inference and Model APIs. ✕ A screenshot of the multimodal serving workflow from Kyle Kranen’s talk at Dynamo Day, “Under the Hood: Dynamo Inference for Multimodal AI Workloads.” KV cache offloading KV cache offloading uses the full memory hierarchy to store KV cache (like CPU, local SSD, or network/remote storage) and bring it back as needed, allowing higher concurrency and context lengths without GPU memory becoming the bottleneck. In Dynamo, KV blocks are managed by the KV Block Manager, and data moves across tiers via NIXL (NVIDIA Inference eXchange Layer), a low-level library used by Dynamo to efficiently move inference state between GPUs, CPUs, and nodes. With disaggregated serving,...