Building Scalable, End-to-End Generative AI with NVIDIA NeMo Framework on Marlowe

Wednesday, October 29, 2025 | 2:00 - 3:30 PM PDT CoDa Building, Rm E401, 4th Floor

Abstract

As foundation models grow in scale and complexity, the need for reproducible, modular, and high-performance research frameworks has become critical. NVIDIA NeMo is an open, extensible framework built on PyTorch and tightly integrated with Megatron-Core and Transformer Engine for optimized distributed training on NVIDIA GPUs, providing a unified stack for developing end-to-end generative AI models.

This talk will introduce key components of the NeMo ecosystem, including data curation, model customization through techniques like parameter-efficient fine-tuning (PEFT) and model alignment algorithms, model safety through guardrails, as well as 4D parallelism for efficient large-model training. Drawing on examples from robotics research, the presentation will highlight how NeMo can be used to build and adapt foundation models for multimodal and domain-specific reasoning, illustrating how NeMo’s modular design enables scalable, domain-specific generative AI for complex systems.

Speaker

Sugandha Sharma is a senior generative AI architect and scientist at NVIDIA, specializing in generative models for robotics and embodied AI. Prior to joining NVIDIA, she was a research scientist at Microsoft Research, where she worked on GPT-based gaming AI agents and their alignment with humans. She holds a PhD in theoretical neuroscience from MIT, where she developed generative AI agents for 3D spatial planning, a memory model without catastrophic forgetting, and the first neural circuit model linking memory and spatial navigation, significantly advancing the theoretical understanding of brain circuits.

Open to Stanford and NVIDIA affiliates.

Part of the ongoing NVIDIA & Marlowe seminar series.