Sri Alavandar Pabbisetty Email & Phone Number

Andhra Pradesh, India

Conducted comprehensive research on Neural Speed Library and Bestla Low precision Kernels.

Optimization of DNN Graphs and Library to enhance CPU performance.- Root cause the performance gap observed in LLMs while using Neural Network library and TensorFlow on specific hardware, and identify the opportunities for performance improvement.- Improved CPU Inference by 52% with LLM variants by implementing Graph Fusions, Library optimizations.- Used OpenMP pragmas to get the best performance during Library optimizations.- Enhanced end to end performance by 40% with Hugging Face LLM models by using optimal Memory allocator (Best-Fit with Coalescing) with TensorFlow.- Achieved 20% improvement with Inception variant by using InterOP parallelism with TensorFlow.- Provided memory pool support with low-precision data type (INT8), resulting in 20% end-end performance.- Added support for specific kernel within TensorFlow when it is integrated with Neural Network Library.- Added support for graph dump feature within TensorFlow when it is integrated with Neural Network Library that helps in model visualization.- Developed validation scripts to ensure model precision.

Chennai, Tamil Nadu, India

- Model conversion from PyTorch to TF through ONNX.- YOLO model conversion from TF (FP32) to TF lite (INT8) format.- Designed an inference pipeline for Person Detection, Pose Estimation and 3D Regressor, which includes the necessary pre- and post-processing steps.- Post Processing: Person Detection (Bounding Boxes - NMS, Threshold) Pose Estimation & 3D Regressor (Temporal smoothing)

Bachelor Of Technology, Computer Science, 8.91

Course Work: Data Structures and Algorithms, Object Oriented Programming, Design and Analysis of Algorithms, Operating Systems, Machine.

Intermediate, Mpc, 96.1%

Ssc, 9