ML Framework (metallm) Engineer

Apple Apple · Big Tech · Cupertino, CA +1 · Software and Services

This role focuses on engineering for ML inference and training frameworks on Apple's custom server hardware and GPUs. The engineer will optimize code for distributed inference, develop kernel and compiler optimizations, apply model optimization techniques, and improve performance metrics. Experience with GPU programming, system-level programming, and distributed training/inference is required.

What you'd actually do

  1. Work on cutting-edge ML inference framework project and optimize code for efficient and scalable ML inference using distributed compute strategies such as data, tensor, pipeline and expert parallelism.
  2. Develop kernel and compiler level optimizations and perform in-depth analysis to ensure the best possible performance across Server hardware families.
  3. Apply advanced model optimization techniques including speculation, quantization, compression, and others to maximize throughput and minimize latency.
  4. Collaborate closely with hardware, compiler, and systems teams to align software performance with hardware capabilities.
  5. Analyze and improve performance metrics such as end-to-end latency, TTFT, TBOT, memory footprint, and compute efficiency.

Skills

Required

  • C/C++/ObjC
  • GPU kernel development
  • Metal
  • CUDA
  • System level programming
  • Computer architecture
  • Distributed training
  • Distributed inference

Nice to have

  • Graph compilers
  • Triton
  • OpenXLA
  • LLVM/MLIR
  • PyTorch
  • JAX
  • Tensorflow
  • Machine learning fundamentals

What the JD emphasized

  • 3+ years of programming and problem-solving experience with C/C++/ObjC
  • Experience with GPU kernel development & optimizations using compute programming models such as Metal, CUDA etc.
  • Experience with system level programming and computer architecture
  • Experience with Distributed training or inference techniques

Other signals

  • enabling Apple Intelligence through high-performance, distributed inference of GenAI applications
  • GPU acceleration of ML Training frameworks
  • scalable, efficient, and production-grade solutions tailored for high-throughput GPU execution
  • optimize code for efficient and scalable ML inference using distributed compute strategies
  • Apply advanced model optimization techniques including speculation, quantization, compression
  • Analyze and improve performance metrics such as end-to-end latency, TTFT, TBOT, memory footprint, and compute efficiency
  • Implement features of Metal device backend for ML training acceleration technologies