Cluster Engineer

STN Inc; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

STN Inc in San Francisco is seeking an experienced AI Infrastructure Engineer to design, deploy, and manage large-scale GPU clusters for AI training and inference workloads.

You will optimize GPU utilization, tune NCCL, CUDA, UCX, and Slurm, and work across storage, networking, and software layers to push performance and scalability. This role requires deep Linux expertise, hands-on container workloads with Pyxis/Enroot, and the ability to implement repeatable benchmarking and automation.

Qualifications

• 7+ years designing or operating large-scale Linux infrastructure.
• 5+ years supporting production GPU clusters for AI or HPC workloads.
• Demonstrated experience building multi-node GPU training environments from the ground up.
• Deep expertise with distributed PyTorch training.
• Extensive experience troubleshooting and optimizing NCCL communications.
• Strong understanding of distributed AI communication patterns: AllReduce, ReduceScatter, AllGather, Broadcast, point-to-point.
• Experience benchmarking distributed training using nccl-tests, NVIDIA DCGM, Nsight Systems, MLPerf (preferred).
• Strong understanding of GPU memory management: KV Cache, activation checkpointing, tensor/pipeline/data parallelism.
• Experience tuning CUDA, NCCL, UCX, and MPI for max distributed performance.
• Expert-level Linux systems administration skills.
• Experience with Slurm workload manager.
• Experience using Pyxis and Enroot for containerized GPU workloads.
• Strong scripting skills in Python and Bash.

Responsibilities

• Design, deploy, and optimize multi-node GPU clusters for AI training and inference workloads.
• Tune distributed training environments to maximize GPU utilization, throughput, and scaling efficiency.
• Optimize inference clusters for maximum token generation throughput, low latency, and high GPU utilization.
• Build and support production AI infrastructure running hundreds to thousands of GPUs.
• Analyze and eliminate performance bottlenecks across compute, networking, storage, and software layers.
• Perform NCCL benchmarking, analysis, and tuning to achieve optimal collective communication performance.
• Design and optimize GPU networking using InfiniBand or RoCE v2, including RDMA, congestion management, topology awareness, and QoS.
• Configure and tune distributed AI software stacks including: PyTorch, NCCL, CUDA, UCX, MPI, Slurm, Pyxis/Enroot.
• Optimize GPU scheduling and resource allocation for both training and inference environments.
• Develop repeatable benchmarking and validation processes for new hardware, firmware, drivers, and software releases.
• Identify performance regressions and troubleshoot distributed training issues at scale.
• Optimize storage architectures for AI workloads, including checkpointing, dataset streaming, and high-performance parallel I/O.
• Work closely with ML engineers to improve training scalability and inference efficiency.
• Create automation to deploy, validate, benchmark, and monitor GPU clusters.
• Evaluate emerging AI infrastructure technologies and recommend improvements to platform architecture.

Skills

GPU clusters

Distributed PyTorch

NCCL

CUDA

UCX

MPI

Slurm

Pyxis/Enroot

Python

Bash

Linux systems administration

Tools

InfiniBand

RoCE v2

RDMA

GPUDirect RDMA

GPUDirect Storage

NVIDIA DCGM

Nsight Systems

nccl-tests

MLPerf

Job description

Position Summary

We are seeking a highly experienced AI Infrastructure Engineer to architect, deploy, optimize, and operate large-scale GPU clusters supporting state-of-the-art AI training and inference workloads. This is a deeply technical role focused on maximizing cluster efficiency, scalability, and performance across the entire AI stack—from GPU hardware and high-speed networking to distributed training frameworks and inference optimization. The ideal candidate has built GPU clusters from the ground up, tuned distributed training environments, optimized large-scale inference deployments, and understands how every layer of the infrastructure contributes to application performance.

Responsibilities

• Design, deploy, and optimize multi-node GPU clusters for AI training and inference workloads.

• Tune distributed training environments to maximize GPU utilization, throughput, and scaling efficiency.

• Optimize inference clusters for maximum token generation throughput, low latency, and high GPU utilization.

• Build and support production AI infrastructure running hundreds to thousands of GPUs.

• Analyze and eliminate performance bottlenecks across compute, networking, storage, and software layers.

• Perform NCCL benchmarking, analysis, and tuning to achieve optimal collective communication performance.

• Design and optimize GPU networking using InfiniBand or RoCE v2, including RDMA, congestion management, topology awareness, and QoS.

• Configure and tune distributed AI software stacks including:

  • PyTorch
  • NCCL
  • CUDA
  • UCX
  • MPI
  • Slurm
  • Pyxis/Enroot

• Optimize GPU scheduling and resource allocation for both training and inference environments.

• Develop repeatable benchmarking and validation processes for new hardware, firmware, drivers, and software releases.

• Identify performance regressions and troubleshoot distributed training issues at scale.

• Optimize storage architectures for AI workloads, including checkpointing, dataset streaming, and high-performance parallel I/O.

• Work closely with ML engineers to improve training scalability and inference efficiency.

• Create automation to deploy, validate, benchmark, and monitor GPU clusters.

• Evaluate emerging AI infrastructure technologies and recommend improvements to platform architecture.

Required Qualifications

• 7+ years designing or operating large-scale Linux infrastructure.
• 5+ years supporting production GPU clusters for AI or HPC workloads.
• Demonstrated experience building multi-node GPU training environments from the ground up.
• Deep expertise with distributed PyTorch training.
• Extensive experience troubleshooting and optimizing NCCL communications.
• Strong understanding of distributed AI communication patterns, including:
  • AllReduce
  • ReduceScatter
  • AllGather
  • Broadcast
  • Point-to-point communications
• Experience benchmarking distributed training using tools such as:
  • nccl-tests
  • NVIDIA DCGM
  • Nsight Systems
  • MLPerf (preferred)
• Strong understanding of GPU memory management, including:
  • KV Cache
  • Activation checkpointing
  • Tensor Parallelism
  • Pipeline Parallelism
  • Data Parallelism
• Experience optimizing LLM inference throughput, including:
  • Tokens/sec optimization
  • Batch sizing
  • Continuous batching
  • KV cache tuning
  • Memory bandwidth optimization
• Experience tuning CUDA, NCCL, UCX, and MPI for maximum distributed performance.
• Expert-level Linux systems administration skills.
• Experience with Slurm workload manager.
• Experience using Pyxis and Enroot for containerized GPU workloads.
• Strong scripting skills using Python and Bash.

Technical Expertise

AI Frameworks

• PyTorch
• CUDA
• NCCL
• Triton (preferred)
• TensorRT-LLM (preferred)

Cluster Scheduling

• Slurm
• Pyxis
• Enroot

GPU Networking

• InfiniBand
• RoCE v2
• RDMA
• GPUDirect RDMA
• GPUDirect Storage
• UCX
• MPI
• Network topology optimization
• Congestion control
• QoS
• ECN/PFC
• High-speed Ethernet (200/400/800 GbE)

Storage

• Parallel file systems
• Distributed storage
• Object storage
• NVMe
• Checkpoint optimization
• Dataset staging
• GPUDirect Storage
• Storage bandwidth optimization
• Metadata performance

Performance Engineering

• NCCL benchmarking
• Multi-node scaling analysis
• GPU utilization optimization
• Communication/computation overlap
• NUMA optimization
• CPU affinity
• PCIe topology
• GPU topology (NVLink/NVSwitch)
• Memory bandwidth analysis
• End-to-end performance profiling

Preferred Qualifications

• Experience deploying AI workloads on Kubernetes.
• Experience with NVIDIA GPU Operator.
• Experience with Kubernetes batch scheduling (Volcano, Kueue, Run:ai, etc.).
• Experience with distributed inference platforms such as vLLM, TensorRT-LLM, or SGLang.
• Experience with NVIDIA DGX SuperPOD or similar large-scale GPU deployments.
• Familiarity with MLPerf benchmarking.
• Experience deploying monitoring solutions such as Prometheus, Grafana, and DCGM Exporter.
• Experience automating infrastructure using Ansible, Terraform, or similar tools.
• Experience working in cloud GPU environments (AWS, Azure, GCP) in addition to bare metal.