Software Engineer, Data Infrastructure

Why this role?

The Data Infrastructure team at Cohere is responsible for the storage and data movement layer underlying every model training run. We're building the unified storage layer that feeds our training workloads. It needs to serve petabytes of training data and model checkpoints fast enough to keep thousands of GPUs busy across several training clusters. In this role, you’d have an opportunity to build this system from the ground up. You’d be a key contributor, working on a problem few teams have had to solve at this scale.

In this role, you will

• Design, build, and operate the distributed storage system that feeds model training and evaluation.
• Run this system multiple on Kubernetes clusters at petabyte scale.
• Work with researchers and training-infra teams on how jobs actually read and write data, and turn that into throughput, latency, and durability requirements
• Work through the networking, I/O, and consistency problems of moving large datasets and checkpoints across regions and backends, with GPU idle time and time-to-insight as the measures of success

You may be a good fit if you have

• Strong storage fundamentals, including replication, consistency, caching, and data lifecycle management.
• Strong coding ability. We work in Python and Go; experience in either is enough, but you should be willing to pick up the other
• Experience running stateful systems on Kubernetes, including Persistent Volumes, CSI drivers, and StatefulSets.
• Hands-on experience with cloud object storage such as S3 as well as POSIX-style filesystems.

It’s a bonus if you have

• Experience with parallel or HPC filesystems such as Weka, VAST, or Lustre.
• Familiarity with the data-loading and checkpointing patterns used in large-scale model training.
• An interest in how large models are trained and evaluated, including the trajectory and eval data those runs generate, and a desire to understand the workloads the platform is serving.