A running list of things I've read and enjoyed — other people's blogs, books, essays. I add to it whenever something earns a spot.
- Look Ma, No Bubbles! — Hazy Research on fusing an entire Llama-1B forward pass into a single GPU kernel, eliminating pipeline bubbles for 1.5x speedup.
- Cascade Inference — FlashInfer's approach to shared-prefix batch decoding, splitting attention between shared and unique KV-cache segments for big throughput gains.
- Radix Top-K — linear-time top-k selection algorithm that buckets elements by bit chunks instead of sorting the whole array.