Senior / Staff ML Engineer, Training Services, Apple Cloud AI
Software Engineering, Data Science
USD 175k-308,500 / year + Equity
Posted on Nov 14, 2026
Apple is a place where extraordinary people gather to do their best work. Apple Cloud AI is the organization that builds Apple's central platform for developing machine learning and generative AI. Researchers, ML engineers, and software engineers across Apple rely on it to train, customize, evaluate, and optimize the models and AI agents behind their features. We provide these capabilities as secure, reliable, managed services on large-scale GPU and TPU infrastructure, so teams can move from an idea to a production-ready model quickly while meeting Apple's high bar for privacy.
Within Apple Cloud AI, our team owns model training services, from pre-training through post-training and distillation, for foundation, recommendation, embedding, speech, and vision models. We're looking for a Staff ML Engineer who understands model architectures and training algorithms from first principles and can translate that knowledge into better models from the same compute.
- Accelerate post-training: Speed up and optimize supervised fine-tuning, parameter-efficient methods such as LoRA and QLoRA, preference optimization such as DPO, and teacher-student distillation, including multi-adapter training and efficient in-loop teacher inference.
- Optimize training performance: Identify and eliminate bottlenecks across computation, memory, communication, and data loading, using profilers, execution traces, and benchmarks to improve throughput and model FLOPs utilization (MFU).
- Advance distributed training: Design and optimize data, tensor, pipeline, context, and expert parallelism, including MoE routing and load balancing, communication collectives, and compute-communication overlap.
- Improve model execution: Develop efficient attention implementations, optimize sequence processing and memory usage, and build compiler optimizations and custom accelerator kernels.
- Improve convergence and stability: Tune mixed-precision execution, quantization-aware training, optimizer configurations, and parameter and optimizer-state sharding while validating model quality.
- Optimize generation-in-the-loop training: Improve generation efficiency, model weight synchronization, and coordinated training-inference execution for workflows where models generate their own training data or signals.
- Make large runs reliable: Build fast checkpointing and resume, fault tolerance, and early detection of hangs and training instability.
- Drive technical direction: Partner with ML researchers on hard training problems, turn one-off fixes into reusable optimizations, and contribute upstream to open-source training frameworks.
- Design training interfaces: Build the APIs and abstractions teams use to train models, from managed end-to-end workflows down to low-level primitives such as forward and backward passes, optimizer steps, and sampling.
- 8+ years of software engineering or ML systems experience, including several years of hands-on distributed training on hundreds of accelerators or more.
- Strong understanding of modern model architectures, especially transformers: model computation, memory requirements, and how architecture choices affect training cost.
- Hands-on experience optimizing compute, memory, and communication in distributed training with PyTorch Distributed, Megatron-Core, JAX/XLA, or equivalent systems.
- Strong programming skills in Python, plus C++, CUDA, Triton, or equivalent performance-oriented technologies.
- Demonstrated ability to independently diagnose complex performance problems, implement optimizations, and validate improvements.
- Proven record of leading technical work across teams and communicating complex tradeoffs clearly.
- BS, MS, or PhD in Computer Science, Machine Learning, or a related field, or equivalent practical experience.
- Deep expertise in mixture-of-experts and expert parallelism, long-context models, or advanced attention mechanisms.
- Experience developing high-performance kernels in CUDA, Triton, or Pallas, or optimizing execution through compilers such as XLA or torch.compile.
- Experience with GPU and TPU architecture, distributed communication, and low-precision training such as FP8.
- Significant contributions to open-source projects such as PyTorch, TorchTitan, Megatron-Core, DeepSpeed, JAX, OpenXLA, AXLearn, MaxText, Triton, FlashAttention, or Transformer Engine.
- Experience designing training APIs or frameworks that other engineers build on, such as trainer abstractions, configuration systems, or low-level training primitives.
- Experience with optimization and training-algorithm research, such as new optimizers, hyperparameter transfer (muP), or scaling-law studies.
- Experience training recommendation, multimodal, or diffusion models at scale.