Speaker
Description
AI workloads on typically involve two distinct usage patterns: scheduled training and fine-tuning jobs, and interactive model inference. Both require access to accelerator resources, but differ substantially in how these resources are allocated and used.
The AI Training App, built on the LEXIS Platform, provides a unified interface for both modes. For training, it simplifies data management, environment configuration and execution of AI workloads on the EuroHPC JU supercomputers, with prepared workflows for models such as YOLO and Whisper, and Unsloth. For inference, the integrated EXA4MIND AI Inference Service enables interactive access to models running on HPC resources through an OpenAI-compatible API, while handling the underlying resource and model management.
The poster presents the architecture and practical use of this unified approach, showing how higher-level interfaces can make both batch-oriented AI training and interactive inference available on HPC infrastructure.