Inspiration

What it does

How I built it

Challenges I ran into

Accomplishments that I'm proud of

We wrote a fully reproducible CI audit that builds llama.cpp (tag b10434) twice — default and -DGGML_CPU_KLEIDIAI=ON with KleidiAI v1.24.0 — on a free ubuntu-24.04-arm GitHub Actions runner (which is a Neoverse-N2, the exact silicon in question). It benchmarks Qwen2.5-1.5B at Q8_0 and Q4_K_M (512-token prompt, 128-token decode, 4 threads, 2 reps), probes svcntb() live, dumps the kernel symbol table, and disassembles SVE vs NEON objects for register-level evidence. Everything is pinned (LLAMA_NATIVE=OFF) so any judge can reproduce the entire audit in ~15 minutes on $0 infrastructure — either via the CI pipeline or bash .github/audit.sh.

What I learned

What's next for Neoverse SVE Dispatch Audit

Built With

Share this project:

Updates