Senior AI Software Architect

Do you want to be at the forefront of innovating the latest hardware designs to propel Microsoft’s cloud growth? Are you seeking a unique career opportunity that combines technical capabilities, cross-team collaboration with business insight and strategy?  

Microsoft’s mission is to empower every person and every organization on the planet to achieve more. As employees, we come together with a growth mindset, innovate to empower others, and collaborate to achieve our shared goals. Each day we build on our values of respect, integrity, and accountability to create a culture of inclusion where everyone can thrive at work and beyond. In alignment with our Microsoft values, we are committed to cultivating an inclusive work environment for all employees to positively impact our culture every day. 

Join the Systems Planning and Architecture (SPARC) team within Microsoft’s Azure Hardware Systems and Infrastructure (AHSI) organization, the team behind Microsoft’s expanding Cloud Infrastructure and for powering Microsoft’s “Intelligent Cloud” mission. Microsoft delivers more than 200 online services to more than one billion individuals worldwide, and AHSI is the team behind our expanding cloud infrastructure. We deliver the core infrastructure and foundational technologies for Microsoft's cloud businesses including Microsoft Azure, Bing, MSN, Office 365, OneDrive, Skype, Teams and Xbox Live.    

We are seeking a highly skilled Senior AI Software Architect to join our team focused on model enablement and performance optimization for Maia accelerators. This role is ideal for someone with strong experience in PyTorch-based model development, quantization techniques, and parallelization strategies at the framework level. You will work closely with hardware and software teams to bring up models on Maia and ensure they run efficiently. 

Model Enablement: 

• Port and optimize large-scale AI models (e.g., foundation models, diffusion models, YOLO) to run efficiently on Maia hardware. 

• Integrate models using frameworks such as PyTorch, ONNX, vLLM, and SGLang. 

Performance Optimization: 

• Apply techniques like KV cache quantization (e.g., BF16 → FP8), checkpointing, and re-sharding for efficient inference and training. 

• Experiment with parallelism strategies (TP, PP) and analyze performance impacts across interconnects (NVLink vs PCIe). 

Inference Stack Development: 

• Collaborate on improving inference pipelines, including KV caching in sglang/vllm and performance tuning at the PyTorch level. 

• Work with Triton kernels for basic operations (e.g., FP8 dequantization) and assist in kernel performance analysis. 

Cross-Team Collaboration: 

• Partner with hardware architects and kernel developers for co-design discussions. 

• Communicate effectively with multiple stakeholders to align on performance goals and deliverables. 

Required Qualifications: