chris_

I work on LLM inference: making models decode faster and fit on less hardware.

Germany/GitHub since 2017/@ultimatechris everywhere

github/ultimatechris huggingface/ultimatechris npm/~ultimatechris

~/ ls ./projects --sort=interesting

Speculative decoding in SGLang using DFlash heads. Ornith-1.5-9B decodes about 4.6× faster, the 35B-A3B MoE about 2.2×. Benchmark setup and reproduction steps are in the repo.

sglangspeculative-decodinginferencebenchmarks

bpmn-di-gennpmv0.1.0 · MIT

Generates readable Diagram Interchange information for BPMN 2.0 XML and lays out swimlanes and flows automatically.

bpmnauto-layoutdiagramstypescript

hf-modelshugging face7 models

Mostly quantizations and finetunes. EXL3 4bpw builds for ExLlamaV3 and TabbyAPI, imatrix GGUF for llama.cpp, and the DFlash heads for the SGLang work.

exl3ggufquantizationmoe

~/hf ls models/ | head -7

Ornith-1.5-35B-A3B-EXL3-4bpwexllamav3 · ↓ 404 Qwen3-Swallow-32B-SFT-v0.2-GGUFllama.cpp · ja/en · ↓ 1.3k Qwen3-Swallow-32B-RL-v0.2-GGUFllama.cpp · ja/en · ↓ 1.1k Ornith-1.5-9B-EXL3-4bpwexllamav3 · ↓ 278 Nemotron-3-Nano-30B-A3B-EXL3exllamav3 · moe Ornith-1.5-9B-DFlash-SGLangsglang · spec-decode heads Ornith-1.5-35B-A3B-DFlash-SGLangsglang · spec-decode heads

~/ cat about.txt

I work on running models, not training them: quantization (EXL3, GGUF), serving stacks (SGLang, ExLlamaV3/TabbyAPI, llama.cpp), and speculative decoding where latency matters. I benchmark everything and put reproduction steps in the repos.