Software Engineer, Workload Enablement
Software Engineer enabling production AI workloads on new hardware platforms through porting, benchmarking, stress testing, and performance optimization. Requires 5+ years in ML systems, distributed training, PyTorch, and RDMA/NCCL expertise.
About the job
Key Responsibilities
- Port and validate key inference and training workloads on new platforms/SKUs as they arrive; drive correctness, performance, and stability to an internal readiness bar.
- Build a suite of benchmarks and stress tests that capture real E2E behavior of our workloads by exercising all aspects of a system, including CPU, GPU, memory subsystem, frontend, scale-up, and scale-out networking (including WAN traffic, NVlink and RDMA collectives), storage, thermals, and any other relevant parts.
- Deep-dive performance on distributed training/inference:
- Collective performance and tuning (across NCCL/RCCL and internal libraries)
- Overlap of compute/communication, kernel-level bottlenecks, memory bandwidth and scheduling effects
- Create repeatable test harnesses that run in CI / lab environments and produce actionable outputs (pass/fail, performance score, regression detection).
- Partner with systems + fleet bring-up engineers to ensure the platform is not only stable and performant, but also operationally usable and scalable (containerization, K8s integration, telemetry hooks, failure triage loops).
- Work cross-functionally with vendors and internal stakeholders by producing clear bug reports, minimal repros, and prioritized issue lists.
Qualifications
- BS in CS/EE (or equivalent practical experience).
- 5+ years in one or more of: ML systems, performance engineering, distributed systems, or HPC.
- Strong hands-on experience with:
- PyTorch and modern LLM training/inference stacks
- Large-scale distributed training concepts (data/model/pipeline parallel, collective comms)
- Experience with RDMA and debugging/optimizing comms libraries (NCCL or RCCL) and their interaction with hardware/network
- Proficiency in Python plus comfort reading/writing performance-critical code (C++/CUDA/HIP is a plus).
- Strong profiling/debugging skills (e.g., Nsight, rocprof, perf, flamegraphs; ability to reason from traces/counters).
Preferred Skills
- Experience building workload-shaped benchmarks and stress/fault tests that correlate to production behavior (not just synthetic loops or microbenchmarks).
- Familiarity with RDMA networking and transport tuning; understanding of how network topology and congestion impact collectives.
- Experience running and validating workloads in Kubernetes, and bridging “research code” into robust, repeatable infrastructure.
- Hands-on lab experience with early hardware (new NICs, new GPUs/accelerators, early racks).
Skills
PyTorch, Nccl, Rccl, Rdma, Kubernetes, Python, C++, CUDA, Hip, Nsight, Perf, Distributed Systems, Ml Systems, Hpc, Llm Training
Similar jobs
ML Engineering jobs
OpenAI
San Francisco, CA
Build and optimize OpenAI’s inference stack for AWS Trainium across high-performance kernels, compilers, runtimes, and model execution. The role requires systems programming and accelerator experience, with opportunities to solve end-to-end performance problems for frontier-scale AI models.
Anthropic
San Francisco, CABuild and deploy LLM-powered tools, agents, and ecosystem infrastructure with life sciences research institutions. The role requires deep scientific or biomedical research experience, production software development expertise, and the ability to translate partner workflows into scalable AI systems.
OpenAI
San Francisco, CA
Build research infrastructure and tooling that enables AI models to design silicon, including reinforcement learning environments, EDA integrations, evaluations, and experiment workflows. The role requires strong software engineering fundamentals and comfort working across research, tooling, and chip-design systems.
OpenAI
San Francisco, CA
Build and optimize the production LLM inference runtime for frontier models on OpenAI’s custom silicon. The role spans scheduling, distributed execution, memory and KV-cache management, performance tooling, and hardware-software co-design.
Hyperbound
San Francisco, CA
Build and operate machine learning models for sales roleplay, scoring, and coaching products, owning the lifecycle from fine-tuning and evaluation through production and on-device deployment. The role emphasizes open-source models, latency and privacy optimization, and rigorous model testing.