Roy Allela

Training vision-language models at scale fails in predictable ways — node crashes mid-epoch, checkpoints corrupt, and GPU clusters sit idle during recovery. In this hands-on workshop, build resilient VLM training pipelines on Amazon SageMaker HyperPod that survive real-world failures.

Amazon SageMaker HyperPod offers advanced training tools to accelerate scalable, reliable, and secure AI model development. In this hands-on session, learn how to orchestrate distributed training using a Slurm-based cluster.

Chaos To Control: Building the Central Nervous System for Enterprise AI With Gateway Architecture - Christian Cania & Boyu Wang, TrueFoundry PROfile Behavioral Embeddings - Xiaojun Fu, Vincent Young, Allen Miyazawa,Jiahao Wu,  Meta PyTorch‑Powered GNNs for Enterprise SSD I/O Acceleration - Zeyuan...

In this builders’ session, get step-by-step implementation experience fine-tuning Llama 3 8B using the QLoRA strategy. Learn how you can fine-tune Llama 3 models at scale using the memory-efficient PyTorch Fully Sharded Data Parallel (FSDP2) library with access to built-in debugging and GPU utilization metrics.

Purge the GIL: Improved Torch.DataLoader - Michal Szolucha & Rostan Tabet, NVIDIA XFormers - Daniel Haziza, Meta AI  TritonCC: AOT Triton Workflow for TorchScript C++ Runtime - Sijia Chen & Huamin Li, Meta The PyTorch 2.0 Inference Story - Angela Yi, Bin Bao, Sheng Qin & Sherlock Huang, Meta Tensor...