AIML - Staff ML Infrastructure Engineer, ML Platform & Technology - Pre-training Infrastructure

Socket.dev; New York, California (NY, MO); On-site; USD 180,000 - 240,000; Full time

Job summary

Apple is seeking a senior ML systems engineer to drive optimization for large-scale foundation model training on TPUs. You will profile and optimize JAX/XLA workloads across compute, memory, and communication, and develop high-performance TPU kernels for core ML operations.

The role requires 6+ years in high-performance ML or distributed systems, strong Python skills, and expertise in distributed systems and performance optimization.

Qualifications

• 6+ years of experience building or optimizing high-performance ML or distributed systems.
• Proficient in Python or other relevant programming languages.
• Strong understanding of distributed systems, parallel computing, and performance optimization.

Responsibilities

• Drive performance optimization for large-scale foundation model training on TPUs, focusing on efficiency, throughput, and scalability.
• Profile and optimize JAX/XLA workloads across compute, memory, communication, and compilation.
• Develop and optimize high-performance TPU kernels for critical ML operations.
• Collaborate with cross-functional engineers to solve large-scale ML training challenges.
• Lead complex technical projects and mentor engineers in areas of your expertise.

Skills

Python

Distributed systems

Performance optimization

Profiling

Education

Bachelor's degree in Computer Science or Engineering

Tools

JAX

XLA

CUDA

Job description

Apple is where individual imaginations gather together, committing to the values that lead to great work. Every new product we build, service we create, or Apple Store experience we deliver is the result of us making each other’s ideas stronger. That happens because every one of us shares a belief that we can make something wonderful and share it with the world, changing lives for the better. It’s the diversity of our people and their thinking that inspires the innovation that runs through everything we do. When we bring everybody in, we can do the best work of our lives. Here, you’ll do more than join something — you’ll add something!

Description

• Drive performance optimization for large-scale foundation model training on TPUs, focusing on efficiency, throughput, and scalability
• Profile and optimize JAX/XLA workloads across compute, memory, communication, and compilation.
• Develop and optimize high-performance TPU kernels for critical ML operations such as attention and Mixture-of-Experts (MoE)
• Optimize distributed training techniques, sharding strategies, and collective communication over TPU interconnects (ICI/Fabric)
• Research and implement new techniques across the JAX, XLA, and TPU stack to improve end-to-end training performance
• Develop performance profiling, benchmarking, and automated tuning capabilities for large-scale training workloads.
• Collaborate with cross-functional engineers to solve large-scale ML training challenges
• Lead complex technical projects and mentor engineers in areas of your expertise
• Cultivate a team centered on collaboration, technical excellence, and innovation

Minimum Qualifications

6+ years of experience building or optimizing high-performance ML or distributed systemsProficient in Python or other relevant programming languagesStrong understanding of distributed systems, parallel computing, and performance optimizationExperience profiling and optimizing compute-, memory-, or communication-intensive workloadsAbility to clearly communicate complex technical problems and collaborate with partners to develop solutionsBachelor's degree in Computer Science, Engineering, or a related field

Preferred Qualifications

Advanced degree in Computer Science, Engineering, or a related fieldExperience with accelerators such as TPU or GPU and understanding of accelerator architecture and performance characteristicsExperience with JAX, XLA, PyTorch or other ML compiler/runtime stacksExperience developing or optimizing accelerator kernels using Pallas, Triton, CUDA, or similar technologiesExperience optimizing large-scale foundation model training and distributed communication