■ Block 01 / research
Research
Foundations other work depends on. Each entry explains what it is, why it is load-bearing, its open problems, and a reviewed timeline of progress.
Foundations other work depends on. Each entry explains what it is, why it is load-bearing, its open problems, and a reviewed timeline of progress.
Storing the transformer key-value cache in fixed-size non-contiguous blocks, with attention kernels that gather through a block table; the memory manager behind high-throughput LLM serving.
A tiled, fused attention kernel that never writes the full score matrix to GPU memory, using an online softmax; versions 1 to 3 track successive GPU generations.
How dense matrix multiplication reaches near-peak throughput on NVIDIA GPUs: hierarchical tiling, tensor-core MMA instructions, asynchronous copies, and the CUTLASS and CuTe abstractions.