llama-server-build

Automated builds of llama.cpp's llama-server for Aegis-AI.

Produces pre-built binaries for all platforms, published as GitHub Releases tagged to match the upstream llama.cpp version (e.g. b8502). New releases are auto-detected weekly (Monday 04:00 UTC).

Build Matrix

Linux x64

Artifact	GPU	SM Targets
llama-server-{ver}-linux-x64-cpu.tar.gz	—	—
llama-server-{ver}-linux-x64-cuda-12.tar.gz	CUDA 12.8	75–120
llama-server-{ver}-linux-x64-cuda-13.tar.gz	CUDA 13.1	75–120
llama-server-{ver}-linux-x64-vulkan.tar.gz	Vulkan	—

Linux ARM64 — Generic

Artifact	GPU
llama-server-{ver}-linux-arm64-cpu.tar.gz	—
llama-server-{ver}-linux-arm64-cuda-12.tar.gz	CUDA 12
llama-server-{ver}-linux-arm64-cuda-13.tar.gz	CUDA 13

⚠️ Note on generic ARM64 builds: These are compiled with -march=armv9-a which enables SVE instructions not available on Cortex-A72/A76/A78AE CPUs (Raspberry Pi 4/5, Jetson Orin, RK3588). If you see signal=SIGILL (exit 132), use the embedded device builds below.

Linux ARM64 — Embedded Devices (no SVE, safe for all boards)

Built with device-appropriate -march flags. CPU backend uses -march=armv8-a on all GPU-accelerated variants since inference runs on CUDA/Vulkan.

Artifact	Device	Accel	CPU flags
llama-server-{ver}-arm64-jetson-orin-cuda-12.tar.gz	Jetson Orin Nano/NX/AGX	CUDA 12	-march=armv8-a
llama-server-{ver}-arm64-jetson-xavier-cuda-11.tar.gz	Jetson Xavier NX/AGX	CUDA 11	-march=armv8-a
llama-server-{ver}-arm64-rpi5-vulkan.tar.gz	Raspberry Pi 5	Vulkan	-march=armv8-a
llama-server-{ver}-arm64-rk3588-vulkan.tar.gz	Rockchip RK3588	Vulkan	-march=armv8-a
llama-server-{ver}-arm64-a76-vulkan.tar.gz	Orange Pi 5, Rock 5B	Vulkan	-march=armv8-a
llama-server-{ver}-arm64-rpi5-cpu.tar.gz	Raspberry Pi 5	CPU	-mcpu=cortex-a76
llama-server-{ver}-arm64-rpi4-cpu.tar.gz	Raspberry Pi 4	CPU	-mcpu=cortex-a72
llama-server-{ver}-arm64-modern-cpu.tar.gz	RPi5, RK3588, Jetson	CPU	-march=armv8.2-a+dotprod
llama-server-{ver}-arm64-safe-cpu.tar.gz	All ARM64 boards	CPU	-march=armv8-a

Windows / macOS

Artifact	GPU
llama-server-{ver}-windows-x64-cuda-12.zip	CUDA 12.4
llama-server-{ver}-windows-x64-cuda-13.zip	CUDA 13.1
llama-server-{ver}-windows-x64-vulkan.zip	Vulkan
llama-server-{ver}-windows-x64-cpu.zip	—
llama-server-{ver}-windows-arm64-cpu.zip	—
llama-server-{ver}-macos-arm64-metal.tar.gz	Metal
llama-server-{ver}-macos-x64-cpu.tar.gz	—

---

Building locally

The scripts/build-embedded.sh script lets you build any variant locally on your device.

Quick SIGILL fix (Jetson / RPi / RK3588)

If you already have a binary installed but it crashes with signal=SIGILL, use patch-cpu-lib to rebuild only libggml-cpu.so with safe flags and swap it in:

git clone https://github.com/SharpAI/llama-server-build.git
cd llama-server-build

# Replace the SVE-crashing libggml-cpu.so in your existing install:
./scripts/build-embedded.sh b8502 patch-cpu-lib \
  ~/.aegis-ai/llama_binaries/b8502/linux-arm64-cuda-12

# Takes ~5 minutes. Verifies automatically on completion.

Full builds by profile

# Single binary for Jetson Orin/Xavier, RPi 5, RK3588 (modern boards, no RPi4):
./scripts/build-embedded.sh b8502 modern-cpu

# Universal binary — all boards including RPi 4:
./scripts/build-embedded.sh b8502 safe-cpu

# Jetson Orin with CUDA (run natively on the device):
./scripts/build-embedded.sh b8502 jetson-orin-cuda

# Raspberry Pi 5 with Vulkan:
./scripts/build-embedded.sh b8502 rpi5-vulkan

# Rockchip RK3588 with Vulkan (Mali-G610):
./scripts/build-embedded.sh b8502 rk3588-vulkan

# Raspberry Pi 4, CPU only:
./scripts/build-embedded.sh b8502 rpi4-cpu

Install output

All builds produce a tarball in ./dist/. Install into Aegis-AI:

VERSION=b8502
PROFILE=jetson-orin-cuda-12  # or rpi5-vulkan, rk3588-vulkan, etc.

mkdir -p ~/.aegis-ai/llama_binaries/${VERSION}/${PROFILE}/
tar -xzf dist/llama-server-${VERSION}-arm64-${PROFILE}.tar.gz \
  --strip-components=1 \
  -C ~/.aegis-ai/llama_binaries/${VERSION}/${PROFILE}/

Cross-compilation (CPU/Vulkan profiles only)

On an x86_64 Linux host, install the aarch64 cross-toolchain and set CROSS_TRIPLE:

sudo apt install gcc-aarch64-linux-gnu g++-aarch64-linux-gnu

CROSS_TRIPLE=aarch64-linux-gnu \
  ./scripts/build-embedded.sh b8502 safe-cpu

CUDA profiles require building natively on the target device.

---

How it works

1. Weekly (Monday 04:00 UTC), the workflow checks the latest llama.cpp release
2. If the repo doesn't have a matching release it automatically builds all variants
3. Binaries are published as a GitHub Release with the same version tag
4. You can also manually trigger a build from the Actions tab

How Aegis-AI uses these builds

Aegis-AI's config/llama-binary-manifest.json contains url_template entries pointing to this repo's releases. The runtime binary manager downloads the appropriate variant when a user installs or upgrades the AI engine.

License

The built binaries are subject to the llama.cpp license (MIT).