Upstream UpdateNullstack contributes Spectrum diffusion & CDNA3 attention upstream to SGLangRead PR log
Nullstackby NullStack
NullStack

Research & Engineering Posts

Research notes, engineering write-ups, and upstream system updates from NullStack.

August 9, 2026
SGLangResearchAMDTPU

Nullstack ♥ SGLang

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.

Read
July 24, 2026
AMDResearchLiteAttention

A Fast Attention Kernel for MI300X, Written in HIP, Not Assembly

How we achieved near-peak memory bandwidth on AMD CDNA3 with clean C++ HIP code, beating hand-tuned assembly through careful memory hierarchy scheduling.

Read
June 18, 2026
ResearchLiteLinear

HyperQuant: Rate-Distortion Optimal Mixed-Precision Quantization

Achieving 3.2-bit effective weights with zero perplexity degradation on 70B+ open models using non-uniform lattice vector quantization.

Read
July 30, 2026
AMDSGLangResearch

MLA Attention for Kimi-K2.7 on AMD MI300X

Benchmarking NullStack MLA against upstream backends on AMD gfx942 CDNA3 hardware with multi-head latent verification.

Read
May 12, 2026
LiteLinearResearch

LiteLinear v2: Ultra-Fast Quantized GEMM for Modern GPUs

W4A16 and W8A16 matrix multiplication kernels designed to bypass dequantization overhead on NVIDIA Hopper and AMD CDNA3.

Read
April 29, 2026
TPUResearch

Zero-Overhead Paged Attention on Google Cloud TPU v5p

Implementing custom XLA HLO fusions to unlock 92% MFU during long-context batch decoding on TPU v5p pods.

Read
April 14, 2026
ResearchSGLang

Optimizing DeepSeek Sparse MLA for 1M Token Contexts

Hardware-aware indexing and cache budgeting for sparse multi-head latent attention kernels.

Read
March 22, 2026
WorldJenPartnership

WorldJen & NullStack: Scaling Open Inference Across Europe

Partnering with European green data center infrastructure to deliver sovereign, GDPR-native AI compute.

Read
February 15, 2026
GMTLVResearch

Generalized Matrix Tensor Lattice Vector Compression (GMTLV)

A mathematical framework for compressing multi-gigabyte weight tensors with bounded Euclidean error guarantees.

Read
January 28, 2026
LiteRunnerResearch

LiteRunner: Microsecond-Scale Distributed Tensor Engine

A C++ runtime designed to orchestrate heterogeneous multi-GPU clusters without Python GIL overhead.

Read
January 10, 2026
BackLiteResearch

BackLite: Just-in-Time Kernel Fusion for Open Inference

Compiling dynamic computational graphs directly into high-throughput device binaries at runtime.

Read