Mostly on post-training, RL environments, and small-model experiments. Write-ups of what I learn while training things.
The first post in a series building up to 5D parallelism. Before we split a model across many GPUs, let's figure out what actually fills a single GPU: parameters, gradients, optimizer states, and activations, which usually end up taking the most.
Why naive KV cache allocation wastes GPU memory, and how PagedAttention fixes it with block-based paging.
10 reward designs, 2 training frameworks, 5 GPU runs. The one-line reward function that finally moved the needle — and why temperature matters more than reward shape.