Reference

Guides

Standalone reference pieces, updated occasionally rather than weekly, for the concepts that keep coming up across issues.

Guide · Reference

Where memory and storage actually live during training

Why training needs roughly 8x the memory footprint of serving the same model, and how that gap is what makes large-scale training a parallelism problem rather than a bigger-GPU problem.

Guide · Reference

Where memory and storage actually live during inference

A visual walkthrough of the data flow from a stored checkpoint to a generated token, and why the memory footprint keeps growing after the model has already loaded.