MLOps Engineer - production GPU

Two halves, and both matter. You make the platform useful — enabling distributed training and production inference on the clusters we build. And you own our infrastructure validation offering: the independent, evidence-based assessment that tells a customer whether the eight-figure GPU estate they just bought actually performs the way the reference architecture promised.

If you like building the benchmark that settles the argument, this is your role.

Workload engineering - Enable and tune distributed training:
PyTorch DDP and FSDP, Deep Speed/ZeRO, tensor and pipeline parallelism, NCCL tuning, MFU measurement and improvement.

- Architect inference serving — Triton, NVIDIA NIM, vLLM, TensorRT-LLM — with quantization, continuous batching, KV-cache and paged attention, and autoscaling to latency SLOs.

- Size inference platforms backward from customer SLOs: time-to-first-token, inter-token latency, concurrency, context length.

- Support the GenAI patterns customers actually ask for — RAG, fine-tuning, agentic pipelines, vector database integration.

Validation offering - Build and own our GPU infrastructure validation suite: fabric validation, NCCL scaling curves, GPU burn and thermal/power soak, storage throughput, GPUDirect verification, reference-workload MFU baselines.

- Deliver acceptance validation engagements and author the signed acceptance reports customers use to hold vendors to their commitments.

- Design the continuous-validation service: post-downtime revalidation, driver and firmware matrix certification, performance drift detection against baseline.

- Produce published benchmark methodology and reports that establish our technical credibility in the market.

- Own PoC workload design and the benchmark evidence that closes engagements.

• 5+ years ML engineering, MLOps, or performance engineering with production GPU workloads.
• Hands‑on distributed training on multi-node GPU clusters — not single‑GPU or notebook-scale work.
• Production LLM inference serving:
  Triton, vLLM, or TensorRT-LLM.
• Rigorous benchmarking discipline — you can profile a system, explain where GPU time actually goes, and defend a number.
• Strong Python; containers and Kubernetes-native workflows (Kubeflow, Ray, Argo Workflows).
• Clear technical writing. A validation report is a deliverable a customer pays for.

NVIDIA NIM and NeMo; MLPerf or formal benchmark program experience; test-and-validation engineering background; fine-tuning at scale (LoRA/QLoRA); customer-facing PoC delivery.