Principal Software Engineer – Large-Scale LLM Memory and Storage Systems
NVIDIA Dynamo 是一个高吞吐量，低延迟推理框架，用于服务于多节点分布式环境中的生成人工智能和推理模型。在性能的 Rust 中构建，为可扩展性而使用 Python,Dynamo 调整了 GPU 片段，路线请求，并在异质集群中管理共享 KV 缓存，使许多加速器感觉像数据中心规模上的单一系统一样。随着大型语言模型迅速超过任何单个 GPU 的内存和计算预算，这个平台能够有效地实现尖端 LLM 工作负载的弹性部署。我们正在寻找一个主要系统工程师来定义大规模 LLM 和存储系统的内存管理的愿景和路线图。你将要做的是：· 设计并发展一种统一的内存层，涵盖 GPU 内存，固定主机内存，RDMA 可访问性内存，SSD 级以及远程文件/对象/云存储以支持大规模 LLm 推断。· 与领先的 LLM 服务引擎 （如 vLLM,SGLang,TensorRT-LLM）进行深度整合并实现集成，重点是 KV 缓存脱载，再使用和在异质和分类群中远程共享。· 共同设计可用于高吞吐量，低延迟推断的接口和协议，使得分类式预填充，同行共享 KV 缓存以及多层次 KV 缓储 （GPU,CPU,本地磁盘和远程内存）可实现。· 与 GPU 架构，网络和平台团队密切合作，利用 GPUDirect,RDMA,NVLink 等类似技术来通过异性加速器和记忆池进行低延迟访问和共享。· 老师和工程师，为记忆和存储子系统设定技术方向，并代表团队在内部审查和年轻论坛中 （开放，会议和面向客户的技术深度潜水）.我们需要看到的是：· 士或博士学位或相等经验 · 15年以上在 C/C++和 Python 中构建大型分布式系统，高性能存储或 ML 系统基础设施的经验，具有生产服务的历史记录。· 深入了解内存层次 （GPU HBM,主机 DRAM,SSD 和远程/对象存储）和设计系统的经验。这些系统涵盖多个级别的性能和成本效率。· 分享缓存或关键值系统，特别是针对低延迟和高同步性优化的设计。·与网络 I/O 和 RDMA/NVMe-oF/NVLink 类技术进行实践经验，并提供了 f
查看英文原文
NVIDIA Dynamo is a high-throughput, low-latency inference framework for serving generative AI and reasoning models across multi-node distributed environments. Built in Rust for performance and Python for extensibility, Dynamo orchestrates GPU shards, routes requests, and manages shared KV cache across heterogeneous clusters so that many accelerators feel like a single system at datacenter scale. As large language models rapidly outgrow the memory and compute budget of any single GPU, this platform enables efficient, resilient deployment of cutting-edge LLM workloads.
We are seeking a Principal Systems Engineer to define the vision and roadmap for memory management of large-scale LLM and storage systems.
What you'll be doing:
- Design and evolve a unified memory layer that spans GPU memory, pinned host memory, RDMA-accessible memory, SSD tiers, and remote file/object/cloud storage to support large-scale LLM inference.
- Architect and implement deep integrations with leading LLM serving engines (such as vLLM, SGLang, TensorRT-LLM), with a focus on KV-cache offload, reuse, and remote sharing across heterogeneous and disaggregated clusters.
- Co-design interfaces and protocols that enable disaggregated prefill, peer-to-peer KV-cache sharing, and multi-tier KV-cache storage (GPU, CPU, local disk, and remote memory) for high-throughput, low-latency inference.
- Partner closely with GPU architecture, networking, and platform teams to exploit GPUDirect, RDMA, NVLink, and similar technologies for low-latency KV-cache access and sharing across heterogeneous accelerators and memory pools.
- Mentor senior and junior engineers, set technical direction for memory and storage subsystems, and represent the team in internal reviews and external forums (open source, conferences, and customer-facing technical deep dives).
What we need to see:
- Masters or PhD or equivalent experience
- 15+ years of experience building large-scale distributed systems, high-performance storage, or ML systems infrastructure in C/C++ and Python, with a track record of delivering production services.
- Deep understanding of memory hierarchies (GPU HBM, host DRAM, SSD, and remote/object storage) and experience designing systems that span multiple tiers for performance and cost efficiency.
- Distributed caching or key-value systems, especially designs optimized for low latency and high concurrency.
- Hands-on experience with networked I/O and RDMA/NVMe-oF/NVLink-style technologies, and familiarity with concepts like disaggregated and aggregated deployments for AI clusters.
- Strong skills in profiling and optimizing systems across CPU, GPU, memory, and network, using metrics to drive architectural decisions and validate improvements in TTFT and throughput.
- Excellent communication skills and prior experience leading cross-functional efforts with research, product, and customer teams.
Ways to stand out from the crowd:
- Prior contributions to open-source LLM serving or systems projects focused on KV-cache optimization, compression, streaming, or reuse.
- Experience designing unified memory or storage layers that expose a single logical KV or object model across GPU, host, SSD, and cloud tiers, especially in enterprise or hyperscale environments.
- Publications or patents in areas such as LLM systems, memory-disaggregated architectures, RDMA/NVLink-based data planes, or KV-cache/CDN-like systems for ML.
With highly competitive salaries and a comprehensive benefits package, NVIDIA is widely considered to be one of the technology world's most desirable employers. We have some of the most forward-thinking and hardworking people in the world working for us and, due to outstanding growth, our special engineering teams are growing fast. If you're a creative and autonomous engineer with a genuine passion for technology, we want to hear from you!
Your base salary will be determined based on your location, experience, and the pay of employees in similar positions. The base salary range is 272,000 USD - 431,250 USD.You will also be eligible for equity and benefits.
Applications for this job will be accepted at least until January 13, 2026.This posting is for an existing vacancy.
NVIDIA uses AI tools in its recruiting processes.
NVIDIA is committed to fostering a diverse work environment and proud to be an equal opportunity employer. As we highly value diversity in our current and future employees, we do not discriminate (including in our hiring and promotion practices) on the basis of race, religion, color, national origin, gender, gender expression, sexual orientation, age, marital status, veteran status, disability status or any other characteristic protected by law.Originally posted on Himalayas