Jul 2024 – Present · 1 yr 10 mos · Hyderabad, Telangana, India · On-site

• Part of the ML Solutions team in AI Group, working on compiler for NPUs.
• Debugging problems in increasing offloading percentage of various models to NPU.
• Analyze vision model architectures and Vision Transformer variants to identify optimizations. Onboard public models to the compiler flow, handling analysis, testing, and validation.
• Develop and maintain evaluation pipelines for segmentation, detection, and NLP models in ONNX. Modify models to create new variants by adding/replacing operators and adjust layer configurations for analysis and testing.
• Regularly monitor and maintain 1500+ models, identify issues for resolution to ensure stable convergence. Generate periodic reports summarizing key issues, gaps, and overall status.

Jan 2024 – Jun 2024 · 5 mos · Hyderabad, Telangana, India · On-site

• Performed post-training quantization with ImageNet calibration on 1000+ models across 50+ architectures using INT8, BF16, and BFP16 formats, exploring diverse configurations to optimize accuracy using Vitis AI Quantizer.
• Boosted quantization and evaluation efficiency by over 80% by automating parallel workflows using AMD’s existing RDI infrastructure.
• Developed automation scripts for ONNX model evaluation, memory analysis, validity checks, special operator extraction, and MAC calculation. Contributed to a tooling project to analyze a given set of models, identify similarities, address compiler support gaps, and prioritize key kernel support and compatibility for client models.