Skip to main content

Multi-GPU Scalability

For labs with multiple GPUs, miStudio partitions work:

  • Inference Partition: One GPU stays dedicated to the base model for interactive steering
  • Training Partition: Remaining GPUs handle SAE training via Celery workers
The Celery/Redis Backbone

GPU jobs can take hours or days. Redis stores the task queue, Celery workers execute tasks. Queue 5 training runs with different L1 settings, close your browser, and check results in the morning — research continues on the backend.