Analyze and improve the full training pipeline, including data loading, preprocessing, memory management, forward and backward passes, optimizer execution, checkpointing, and experiment recovery.
Identify bottlenecks across compute, memory, storage, networking, and interconnects to increase accelerator utilization and researcher productivity.
GPU Kernel and Framework Development
Develop and optimize GPU kernels and performance-critical framework components for quantitative machine learning workloads.
Integrate specialized libraries, compilers, and execution techniques to improve throughput, memory efficiency, and numerical performance.
Model and Numerical Optimization
Apply techniques such as mixed-precision training, gradient accumulation, activation checkpointing, operator fusion, and memory-efficient optimizers.
Evaluate tradeoffs among training speed, numerical stability, reproducibility, model quality, and infrastructure cost.
Training Infrastructure
Partner with HPC and infrastructure teams to optimize workload scheduling, resource allocation, observability, fault tolerance, and reproducibility across shared compute environments.
Help define the architecture and tooling required to support large-scale experimentation across on-premises and cloud-based infrastructure.
Cross-Functional Collaboration
Work closely with ML Researchers, Quantitative Researchers, HPC Engineers, Systems Engineers, and hardware specialists to translate research requirements into highly efficient training systems.
Qualifications
3+ years of experience optimizing machine learning training workloads in high-performance, distributed, or large-scale computing environments.
Deep knowledge of machine learning frameworks such as PyTorch or JAX, including their execution models, compilation paths, autograd systems, and distributed-training capabilities.
Strong programming skills in Python and C++, with experience developing or optimizing performance-critical systems.
Proven experience with GPU kernel development and optimization using technologies such as CUDA, Triton, CUTLASS, cuBLAS, cuDNN, or related libraries.
Strong understanding of GPU architecture, including streaming multiprocessor execution, warp scheduling, tensor cores, and the memory hierarchy from registers through HBM.
Experience with distributed-training technologies and communication libraries such as NCCL, FSDP, DeepSpeed, Megatron-LM, XLA, or equivalent systems.
Proficiency with performance-analysis tools such as Nsight Systems, Nsight Compute, PyTorch Profiler, or comparable tracing and profiling platforms.
Understanding of high-performance networking, storage, and accelerator interconnects, including technologies such as InfiniBand, RDMA, NVLink, or NVSwitch.
Demonstrated ability to benchmark heterogeneous compute platforms and make rigorous, data-driven recommendations about performance, scalability, and cost.
Preferred Qualifications
Experience optimizing training workloads for transformer-based, time-series, reinforcement-learning, or other computationally intensive models.
Experience with cluster orchestration and scheduling technologies such as Kubernetes, Slurm, Ray, or similar platforms.
Sponsorship data shows employers' past filings in public US government records (DOL, USCIS, E-Verify) and is not a promise that any job will be sponsored. Immigration rules change. Verify your situation with your DSO or an immigration attorney.