LUMI AI Factory Releases Comprehensive AI-HPC Guide Collection: A Complete Practical Manual from Container Deployment to Quantization Inference

The AI Factory team of Europe's LUMI Supercomputing Center has released an open-source AI-HPC Guide Collection. This collection systematically organizes best practices for running AI workloads on large-scale HPC clusters, covering key topics such as PyTorch containerization, multi-GPU training, LLM fine-tuning, and inference optimization. It serves not only LUMI users but also provides valuable references for AI applications in other HPC centers.

With the rapid development of large language models (LLM) and generative AI, researchers and developers face challenges running complex AI tasks on HPC clusters due to differences in parallel file systems (like Lustre), multi-node GPU communication, software stacks, and scheduling systems. The AI-HPC Guide Collection was launched to address this learning curve.

LUMI is one of Europe's most powerful supercomputers (located in Finland, funded by EuroHPC JU) with AMD Instinct MI250X GPUs and EPYC CPUs, leading in green energy use. Its AI Factory is a dedicated partition for AI/ML workloads, offering optimized hardware and software environments.

The guide is organized by the typical lifecycle of AI workloads on HPC, covering:

1. AI container & software environment configuration (Singularity/Apptainer for PyTorch, file system binding, GPU visibility)
2. Lustre file system data management (format choices like HDF5/Zarr/WebDataset, I/O optimization)
3. LLM fine-tuning (HuggingFace Accelerate, DeepSpeed, Megatron-Bridge, Nanotron examples)
4. Multi-GPU/multi-node training (PyTorch DDP, DeepSpeed configs, MPI/NCCL collaboration)
5. Performance analysis (ROCm-SMI, PyTorch Profiler, hyperparameter optimization)
6. MLOps (TensorBoard, MLflow setup)
7. Inference optimization (vLLM/Ollama, AWQ/BitsAndBytes/GPTQ quantization)
8. Model evaluation (LM Evaluation Harness adaptation)

The guide has three main highlights:

1. Multi-platform adaptation: Applicable to other AMD GPU-based HPC centers (e.g., Finland's Mahti, Italy's Leonardo)
2. Community-driven: Open-source, accepting community contributions via Issues/PRs
3. Practical orientation: Emphasizes runnable code snippets and config files instead of conceptual descriptions, lowering entry barriers

The guide is suitable for:

• AI researchers: Scholars/grad students running large-scale model training on supercomputers
• HPC admins: Optimizing cluster configurations for AI workloads
• MLOps engineers: Deploying/managing AI services on HPC
• AI infrastructure developers: Working on distributed training frameworks or inference engines

Precautions: Most referenced code repositories are not maintained by LUMI AI Factory; users should assess risks independently.

Suggestions: For new HPC AI developers, learn in order: container configuration → single GPU training → multi-node distributed training. Utilize HPC center technical support when encountering issues.

The AI-HPC Guide Collection bridges AI and HPC fields. Its value grows as AI models scale and HPC architectures evolve. Future plans: Expand to cover more hardware platforms (e.g., AMD MI300X, Intel Ponte Vecchio) and software stacks, becoming a shared knowledge base for the global AI-HPC community.