Optimizing deep learning models to enable transformative user experiences on Apple devices by compressing and speeding up models, and productizing model optimization algorithms
Requirements
- Highly proficient in Python programming
- Expertise in shell programming, experience with setting up and/or maintaining CI pipelines for at least one production software codebase
- Proficiency in at least one ML authoring framework, such as PyTorch, TensorFlow, JAX, MLX
- Experience in training, fine tuning, and optimizing neural network models
- Experience in the area of model compression and quantization techniques, specially in one of the optimization libraries for an ML framework (e.g. torch.ao)
- Demonstrated ability to design user friendly and maintainable APIs
- Experience with PyTorch
Responsibilities
- Setting up, and/or streamlining CI and automation pipelines
- Making enhancements to the release process, automating nightly builds, and setting up scheduled CI runs for different levels of testing
- Making innovations in model testing and benchmarking (accuracy and latency), for various combinations of model types in different domains (vision, text, audio etc) and compression algorithms (quantization, pruning, palettization etc)
- Finding innovative ways to reduce test time while maintaining high quality test coverage
- Developing integration of the model optimization library with other training engines and data platforms at Apple
- Keeping the code base updated to work with the latest versions of Python, PyTorch, numpy etc
- Set up and debug training jobs, datasets, evaluation, performance benchmarking pipelines
Other
- Bachelors in Computer Sciences, Engineering, or related discipline
- 2 years of industry experience (including internships)
- Good communication skills, including ability to communicate with cross-functional audiences
- Self prioritize and adjust to changing priorities and asks
- Passionate about the user experience and ways to improve it, to fix bugs, understand user pain points and actively participate in supporting users
