Research notes, engineering write-ups, and upstream system updates from NullStack.
Nullstack contributes performance work upstream to SGLang, because an open PR is only the start: the work matters when it is accepted into the library and usable by everyone.
How we achieved near-peak memory bandwidth on AMD CDNA3 with clean C++ HIP code, beating hand-tuned assembly through careful memory hierarchy scheduling.
Achieving 3.2-bit effective weights with zero perplexity degradation on 70B+ open models using non-uniform lattice vector quantization.
Benchmarking NullStack MLA against upstream backends on AMD gfx942 CDNA3 hardware with multi-head latent verification.
W4A16 and W8A16 matrix multiplication kernels designed to bypass dequantization overhead on NVIDIA Hopper and AMD CDNA3.
Implementing custom XLA HLO fusions to unlock 92% MFU during long-context batch decoding on TPU v5p pods.
Hardware-aware indexing and cache budgeting for sparse multi-head latent attention kernels.
Partnering with European green data center infrastructure to deliver sovereign, GDPR-native AI compute.
A mathematical framework for compressing multi-gigabyte weight tensors with bounded Euclidean error guarantees.
A C++ runtime designed to orchestrate heterogeneous multi-GPU clusters without Python GIL overhead.
Compiling dynamic computational graphs directly into high-throughput device binaries at runtime.