Principal Software Engineer Dynamo

NVIDIA Dynamo is an innovative, open-source platform focused on efficient, scalable inference for large language and reasoning models in distributed GPU environments. By bringing to bear sophisticated techniques in serving architecture, GPU resource management, and intelligent request handling, Dynamo achieves high-performance AI inference for demanding applications. Our team is addressing the most challenging issues in distributed AI infrastructure, and we're searching for engineers enthusiastic about building the next generation of scalable AI systems. As a Principal Software Engineer on the Dynamo project, you will address some of the most sophisticated and high-impact challenges in distributed inference, including: Dynamo k8s Serving Platform: Build the Kubernetes deployment and workload management stack for Dynamo to facilitate inference deployments at scale. Identify bottlenecks and apply optimization techniques to fully use hardware capacity. Scalability & Reliability: Develop robust, production-grade inference workload management systems that scale from a handful to thousands of GPUs, supporting a variety of LLM frameworks (e.g., TensorRT-LLM, vLLM, SGLang). Disaggregated Serving: Architect and optimize the separation of prefill (context ingestion) and decode (token generation) phases across distinct GPU clusters to improve throughput and resource utilization. Contribute to embedding disaggregation for multi-modal models (Vision-Language models, Audio Language Models, Video Language Models). Dynamic GPU Scheduling: Develop and refine Planner algorithms for real-time allocation and rebalancing of GPU resources based on fluctuating workloads and system bottlenecks, ensuring peak performance at scale. Intelligent Routing: Enhance the smart routing system to efficiently direct inference requests to GPU worker replicas with relevant KV cache data, minimizing re-computation and latency for sophisticated, multi-step reasoning tasks. Distributed KV Cache Management: Innovate in the management and transfer of large KV caches across heterogeneous memory and storage hierarchies, using the NVIDIA Optimized Transfer Library (NIXL) for low-latency, cost-effective data movement. What you'll be doing: Collaborate on the design and development of the Dynamo Kubernetes stack. Introduce new features to the Dynamo Python SDK and Dynamo Rust Runtime Core Library. Design, implement, and optimize distributed inference components in Rust and Python. Contribute to the development of disaggregated serving for Dynamo-supported inference engines (vLLM, SGLang, TRT-LLM, llama.cpp, mistral.rs). Improve intelligent routing and KV-cache management subsystems. Contribute to open-source repositories, participate in code reviews, and assist with issue triage on GitHub. Work...