AMARBARO
Home
Skills
Web
Apps
Libs
Blog
Contact
root :: amarbaro.com
EN
DE
FR
NL
ES
IT
PL
RO
◆
projects
Projects
Mojo apps and libraries from amarbaro.org labs. Every page links to the public repository.
corewait: why a CPU instruction stalls, on a desktop or a phone
A function profile says which code is hot. corewait says why each instruction is slow, on a Ryzen desktop, an old Skylake laptop, and three phones, and it can gate a kernel in CI.
ctxd: a context-folding proxy for LLM APIs
A context-folding proxy. Old turns become summaries once a request passes 80% of its window, and the original messages stay readable from a local store.
heaptide: a leak finder for CPU, GPU and Python programs
One run of heaptide gives one report: what was never freed, whether memory keeps rising, who holds it, and which sites allocate most. It sees HIP device and pinned host memory on ROCm.
mojo-baro: an LLM inference engine for AMD RDNA3 GPUs, in Mojo
An inference engine for consumer AMD cards with every kernel in Mojo. On matched Q4_0 weights it decodes 1.34x faster than llama.cpp on the same box, measured on Mojo 1.1.
mojo-neuron: reading and steering the neurons of a language model
It reads one layer of a 0.5B model in 11 seconds on a CPU and steers neurons 4.3 times faster than gguf-neuron-explorer. Most of the research built on top missed its bar, and the README lists those misses beside the passes.
mokei: a line counter in Mojo for 40 languages
A line counter in Mojo that matches tokei on the corpora it was checked against, and adds diff budgets, token estimates and duplicate detection.
wavewait: where a GPU kernel waits, instruction by instruction
A kernel can look busy and still wait. wavewait turns a hardware thread trace into a per-instruction stall table, with a cause on every row, and a budget gate for CI.
mojo-libs: the amarbaro Mojo libraries in one table
Thirty-eight Mojo libraries, each in its own public repository, each with tests against an oracle and a conda package.