Staff Machine Learning Engineer, Apple Cloud AI Platform

Apple Cloud AI is at the forefront of ML and generative AI where data and models come together to enable the next generation of intelligent experiences on Apple products and services. We build the scalable model training services and frameworks that accelerate the development of real-world AI applications designed with both usability and efficiency in mind. Our platform covers the full model training lifecycle from the first experiment to large-scale pre-training fine-tuning and reinforcement learning on thousands of GPUs and TPUs. Our team focuses on applying the latest AI research to training with user experience and accelerator efficiency in mind. Youll play a key role in shaping AI research at Apple and how its models are built optimized and scaled. Were looking for engineers and researchers who are passionate about research and building a high-performance model training stack that supports diverse real-world use cases. With the autonomy to experiment the scale to make an impact and the support to take ideas from prototype to production. Youll work alongside a world-class team to build models and training services that make ML development faster more reliable and more efficient.n

The Apple Cloud AI Platform team works with Apples researchers and ML engineers to train and ship models at Apples bar for quality and this role you will bring deep hands-on experience in large-scale model training to Apples most important training efforts. You will develop training recipes for pre-training fine-tuning reinforcement learning and model distillation and apply techniques such as parameter-efficient fine-tuning preference optimization quantization-aware training. You will diagnose loss spikes and numerical instability and choose the precision parallelism and optimizer settings that let models converge reliably at scale on GPUs and TPUs. As a Staff ML/Research Engineer you will set the technical direction for how Apple trains its models and lead efforts that span multiple teams.n

Work with researchers and ML engineers to develop cutting-edge ML models and turn their needs into training features and performance improvements that every team on the platform can model distillation and compression methods such as teacher-student distillation and quantization-aware training that bring the quality of large models to smaller models with long-context extension mixture-of-experts architectures and multimodal inputs and solve the stability and efficiency problems each one into training behavior at scale such as loss spikes divergence and drops in throughput and track each problem down to its root cause in the data the numerics the code or the and own scalable efficient training frameworks and services on GPUs and TPUs built on PyTorch or JAX and on higher-level reinforcement learning infrastructure for post-training including rollout and inference services reward pipelines and secure sandboxes where models can run code and use the training platform with Apples data and governance stack so that every training run meets privacy and security technical direction across teams mentor engineers and raise the engineering bar.

8 years of experience in large-scale model training including reinforcement learning from research to production.n5 years of experience building and operating large-scale model training systems in with common fine-tuning and distillation techniques such as supervised fine-tuning parameter-efficient fine-tuning with LoRA and teacher-student -on experience with at least one higher-level training framework such as Megatron-LM DeepSpeed TorchTitan NeMo AXLearn or expertise in PyTorch or JAX including distributed training programming skills in Python or collaboration and written and verbal communication skills with a track record of leading technical work across MS or PhD degree in Computer Science Computer Engineering or equivalent practical experience.

Experience with low-precision training and the numerical issues it training large models with data tensor pipeline or expert parallelism across hundreds to thousands of GPUs or understanding of GPU and TPU architecture including memory hierarchy matrix units and high-bandwidth memory and how they shape training with collective communication libraries such as NCCL and with high-speed interconnects such as NVLink InfiniBand and TPU profiling and optimizing training workloads with tools such as Nsight Systems PyTorch Profiler or building reinforcement learning or post-training systems for large source contributions to training frameworks or publications at top ML venues such as NeurIPS ICML or ICLR.