14 hours ago
Base Salary
$102k - $219k/yr
Responsibilities
- Design and develop quantization and numerics algorithms for efficient LLM training and inference in Azure production environments.
- Develop proof-of-concept software and model-optimization tooling to streamline deployment of quantized models.
- Analyze performance bottlenecks in quantized LLM architectures and drive performance improvements.
- Prototype and evaluate low-precision data formats on novel hardware accelerator SDKs.
- Co-design model architectures optimized for low-precision deployment.
- Collaborate with AI/ML researchers, data scientists, hardware architects, and AI software framework teams to align model accuracy and system performance goals.
Requirements
- Bachelor's degree in a related field and 2+ years of related experience, master's degree and 1+ year of related experience, doctorate, or equivalent experience.
- Ability to pass Microsoft's required Cloud Background Check and other applicable security screenings.
- Experience delivering production-grade software for model compression, low-precision numerics, low-level kernel development, and performance optimization.
- Proficiency with PyTorch, TensorFlow, TensorRT, and ONNX Runtime.
- Experience developing GPU/NPU kernels with CUDA, Triton, ROCm, or comparable frameworks.
- Strong understanding of Transformer and LLM architectures and optimization methods including quantization, pruning, sharding, model parallelism, KV-cache optimization, and Flash Attention.
- Experience with large-scale model evaluation, LLM benchmarking, and fine-tuning through SFT or RL.
- Strong programming skills in Python, C, and C++.
- Preferred qualifications include 1+ year of publication experience, low-level linear algebra or custom BLAS kernel experience, and knowledge of mixed-precision arithmetic units, numerical formats, and microarchitecture.
- Strong written, verbal, communication, and cross-team collaboration skills.
Categories
About Microsoft
Microsoft develops operating systems, productivity software, cloud services, developer tools, and consumer devices for individuals, enterprises, and governments. Its main products include Windows, Microsoft 365, Azure, Visual Studio/GitHub, Xbox, and LinkedIn; revenue comes from software subscriptions and licenses, cloud consumption, hardware sales, and advertising. Founded in 1975 and headquartered in Redmond, Washington, Microsoft is a public company traded on Nasdaq.