Cerin Amroth

Independent systems and security research built around inspectable claims: running large models on hardware people actually own, and finding the gaps between what software promises and what it does.

ML systems — quantized Mixture-of-Experts training and inference across GPU memory, host memory, and NVMe

experts4bit-qlora Train and serve Mixture-of-Experts models that do not fit in VRAM: fused 4-bit experts, QLoRA, CPU/NVMe offload, fast inference on consumer NVIDIA GPUs. PyPI experts4bit-qlora · claims tiered in the public register: 161 measured · 6 measured-private · 3 open · 7 superseded · 6 retired grouped-nf4-gemm Triton kernels for 4-bit MoE inference: grouped NF4/MXFP4 GEMM, INT4 GEMV, FP8 paged attention, CPU/NVMe expert streaming. PyPI grouped-nf4-gemm · claims tiered in the public register: 11 confirmed · 9 measured · 3 measured-private · 1 projected · 3 open · 1 superseded · 2 retired OpenVINO Upstream fixes keeping Intel-GPU LLM inference off the CPU-fallback path — three merged.

GitHub: experts4bit-qlora · PyPI · problems it solves · GitHub: grouped-nf4-gemm · PyPI · problems it solves

Start with the problem

Problem	Route
My MoE still OOMs after `load_in_4bit=True`	experts4bit-qlora
I need to QLoRA-train fused MoE experts	experts4bit-qlora
My model is larger than GPU or host memory	e4b residency paths
Dequantizing every expert before GEMM is too slow	grouped-nf4-gemm
I need native MXFP4 or INT4 expert compute	grouped-nf4-gemm
I want to inspect the evidence methodology	Receipts-driven engineering

Observational — JWST imaging, measured rather than benchmarked

Little red dots — three sky maps Where they are: everything published from JWST, everything published without it, and what an independent search found across 89 fields.

Security research — good-faith vulnerability research and coordinated disclosure

Vulnerability research Good-faith hunting & coordinated disclosure.

Contact: jordan@cerinamroth.com · security reports: security@cerinamroth.com · security.txt · consulting: jordananderson.work