https://aaronbrooks.me/https://aaronbrooks.me/about/https://aaronbrooks.me/posts/https://aaronbrooks.me/posts/checkpointing-is-a-distributed-systems-problem/https://aaronbrooks.me/posts/cuda-optimization-starts-with-memory-transactions/https://aaronbrooks.me/posts/i-was-wrong-to-sleep-on-jax/https://aaronbrooks.me/posts/llm-platforms-need-operational-memory/https://aaronbrooks.me/posts/occupancy-is-usually-not-the-first-problem/https://aaronbrooks.me/posts/prompts-are-release-artifacts/https://aaronbrooks.me/posts/training-infrastructure-is-failure-accounting/https://aaronbrooks.me/tags/cuda/https://aaronbrooks.me/tags/distributed-systems/https://aaronbrooks.me/tags/gpu/https://aaronbrooks.me/tags/infrastructure/https://aaronbrooks.me/tags/jax/https://aaronbrooks.me/tags/llm-systems/https://aaronbrooks.me/tags/performance/https://aaronbrooks.me/tags/platforms/https://aaronbrooks.me/tags/production/https://aaronbrooks.me/tags/pytorch/https://aaronbrooks.me/tags/training/