Linear Group systems research

Inference, systems, and chips

Research notes on low-level systems and GPUs: KV cache management, quantization, PCIe, attention kernels, and disaggregated serving. Implementations, measurements, and implications.

John Rose — ex-IBM Linux kernel (author of rpadlpar, PCI hotplug driver for POWER; co-author librtas) · github.com/jnros · email

Research notes

2026-06 DiffusionBlocks AR: first language-model port Block-wise training with GPU decoupling. code
2026-05 Roofline & ridge point: GPU inference optimization Decode, roofline, and ridge-point benchmarks; H100 characterization found <30% of peak memory bandwidth utilized. code
2026-05 KV Mix: asymmetric K/V quantization Outlier-driven mixed precision; K and V tolerate different bit-widths. code
2026-04 DeepSeek V4: system design & implications Analysis of the methods and their systems impact. analysis
2026-04 Rootless agent sandbox from kernel primitives Isolation composed from namespaces, Landlock, and seccomp — no root, no containers. code
2026-03 KV compression (TurboQuant): memory & demand implications What KV compression does to serving capacity per GPU. analysis
2026-03 Paged attention: KV cache as virtual memory (CUDA) From-scratch CUDA implementation treating KV cache as paged virtual memory. code
2026-02 Decode attention: MHA / GQA / MQA Benchmarked attention variants across memory- and compute-bound regimes. code
2026-01 Kimi K2.5: methods + systems impact Analysis of training methods and inference-side consequences. analysis
2026-01 FP64 on Blackwell INT8 tensor cores (Ozaki scheme): POC Proof of concept for emulated double precision on INT8 hardware. code analysis
2026-01 KV offload (Engram): system design Design for KV cache offload across memory tiers. analysis

Background

Seven years in Linux kernel at IBM: architected I/O virtualization for POWER (rpadlpar, still in mainline), created librtas, led the international virtualization team, upstreamed across 12 kernel releases. Fifteen years of technology strategy and capital allocation at AT&T. Now building inference, GPU, and low-level systems full time.