Senior Software Engineer, AI Runtime

At Databricks, we are passionate about enabling data teams to solve the world's toughest problems: from making the next mode of transportation a reality to accelerating the development of medical breakthroughs.

We do this by building and running the world's best data and AI infrastructure platform so our customers can use deep data insights to improve their business.

Training and customizing state-of-the-art AI models is one of the most demanding workloads in computing, and it sits at the heart of Databricks' Mosaic AI mission.

AI Runtime (AIR) is our managed platform for large-scale GPU training and fine-tuning.

It gives customers on-demand access to fleets of the latest accelerators and a serverless experience that hides the complexity of provisioning, scheduling, and orchestrating multi-node jobs, with the resilience to keep training running for days or weeks across thousands of GPUs.

AIR powers the full spectrum of custom training, from fine-tuning open models to pre‑training frontier‑scale foundation models, for some of the most sophisticated AI teams in the world.

Responsibilities

• Build and scale the systems that make large-scale training fast, reliable, and effortless.
• Drive the architecture and evolution of the managed GPU training stack, spanning scheduling and capacity, distributed training performance, fault tolerance, and the developer experience of launching and operating jobs at scale.
• Help shape the technical direction for AIR, mentor other engineers, partner across product, research, and platform teams, and contribute to the initiatives that expand the technical and business impact of custom training at Databricks.
• Deliver scalable, high-throughput, and resilient training across fleets spanning thousands of accelerators.
• Solve the hardest problems in large-scale training, including multi‑node orchestration, distributed parallelism strategies, GPU scheduling and dynamic routing, high‑throughput data loading, and checkpointing and restore for very long-running jobs.
• Push GPU efficiency and training performance, raising utilization (such as model FLOPs utilization and end-to-end throughput) and lowering cost per training run across diverse model architectures and hardware generations.
• Build the resilience and observability foundations that keep multi-node jobs healthy, detecting and recovering from hardware and software failures with minimal disruption to customers.
• Partner with product, research, and platform teams to shape the APIs, CLI, and developer experience that make it easy to launch, monitor, and debug production training jobs.
• Lead end-to-end engineering efforts, from design through production rollout, holding a high bar for performance, correctness, and reliability.
• Make direct, high-impact contributions to the core systems behind AIR, and help bring up support for the latest accelerators and new regions as the fleet grows.
• Champion engineering excellence, mentor other engineers through design reviews and technical discussions, and contribute to Databricks' technical direction in AI training infrastructure.

Qualifications

• 5+ years of experience building and operating large-scale distributed systems, with experience in GPU training infrastructure, high-performance computing, or ML systems.
• Experience with distributed training frameworks such as Py Torch, FSDP, Deep Speed, or Megatron and the parallelism strategies (data, tensor, pipeline, and sequence parallelism) used to train large models.
• Strong understanding of training resilience patterns, including checkpointing, failure detection, and automatic recovery for long-running, multi-node jobs.
• Solid grasp of GPU performance fundamentals, including accelerator architecture, high-speed interconnects such as NVLink and Infini Band or Ro CE, collective communication, and the bottlenecks that govern training throughput and utilization.
• Experience building and operating managed, multi-tenant platform products in the cloud, with clear SLAs and SLOs for availability, performance, and reliability.
• Strong foundation in algorithms, data structures, and system design as applied to performance-sensitive, large-scale distributed systems.
• #J-18808-Ljbffr