higgsfield
GPU cluster orchestration and training framework for very large models
What it is
A Python framework and workload manager for training large neural networks, such as LLaMA-scale language models, across multiple GPU nodes. It allocates nodes to users, queues experiments, supports DeepSpeed ZeRO-3 and PyTorch FSDP sharding, and deploys runs through GitHub Actions workflows. ML engineers running their own Ubuntu GPU servers or rented cloud nodes use it to launch, monitor, and checkpoint distributed training jobs without hand-managing environments and configs.