Infrastructure SRE - HPC

Sarvam runs a large, multi-vendor GPU fleet that serves two demanding workloads on the same physical infrastructure: training jobs that span hundreds of GPUs and must run uninterrupted for weeks, and inference services that must hold a flat p99 under production load. Keeping both healthy at once is a hard, specialized reliability problem, and it is the problem this team exists to solve.

This is not a Kubernetes administration role. We assume Kubernetes fluency as a baseline. The difficulty lies above and below it - in parallel filesystems under heavy checkpoint load, in RDMA fabrics that degrade quietly, in NCCL hangs whose root cause may be the network or the kernel, in driver and firmware drift across heterogeneous hardware, and in distributed training failures that masquerade as infrastructure faults.