Data-parallel training: gradient bucketing and overlap
Why DDP feels like magic until you look at the allreduce schedule.
Ml
A minimal, reproducible Docker setup for ML experiments
Pin everything, mount data, never bake it. The boring setup that stopped me losing runs.
KV cache, and why LLM inference is memory-bound
The cache that makes autoregressive decoding fast also makes it the thing that runs out of memory first.