Principal ML Performance Engineer - GPU Optimization

Sponsor Finder; Boston; Presencial; GBP 113.000 - 159.000; Jornada completa

Descripción de la vacante

Proxima is seeking a Principal ML Performance Engineer to optimize training and inference for structural and generative models, with a focus on CUDA, Triton, and distributed training across large GPU clusters. You will profile bottlenecks, write efficient kernels, and drive multi-node scalability.

As a senior leader, you will scale training across 32–64 nodes, reduce inference costs on thousands of GPUs, and mentor engineers while defining technical direction for the team.

Formación

• 6+ years of experience in ML systems, HPC, or performance engineering.
• Ability to set technical direction beyond coding and mentor engineers.
• Deep knowledge of PyTorch internals with profiling experience.
• Experience with CUDA and Triton and reading Nsight outputs.
• Experience with distributed training at multi-node scale.
• Strong Python and C++ proficiency.
• Ability to name a model they made faster and quantify the improvement.

Responsabilidades

• Profile and optimize training and inference for models including transformers, diffusion, and geometric deep learning.
• Write and tune custom kernels (CUDA, Triton) and use compilers (torch.compile, TensorRT, XLA).
• Scale distributed training across 32-64 nodes with FSDP, DeepSpeed, and mixed precision.
• Reduce inference cost by memory optimization for large complexes and batching inputs.
• Manage GPU cluster efficiency on GCP with scheduling and cost reporting.
• Develop benchmarks and profiling tools for the research team.

Conocimientos

PyTorch internals

CUDA

Triton

Python

C++

Distributed training

Educación

BS/MS/PhD in CS, EE, or related field

Herramientas

Nsight

TensorRT

Torch.compile

XLA

GCP

Descripción del empleo

Proxima is seeking a Principal ML Performance Engineer to optimize training and inference for structural and generative models, with a focus on CUDA, Triton, and distributed training across large GPU clusters. You will profile bottlenecks, write efficient kernels, and drive multi-node scalability.

As a senior leader, you will scale training across 32–64 nodes, reduce inference costs on thousands of GPUs, and mentor engineers while defining technical direction for the team.