Implementation:Ggml org Llama cpp CI Run

Knowledge Sources	Ggml_org_Llama_cpp
Domains	CI, Testing
Last Updated	2026-02-15 00:00 GMT

Overview

Main CI test runner script that builds llama.cpp with various hardware backends and runs the full test suite.

Description

This shell script is the central CI automation entry point for llama.cpp. It accepts output and mount directories as arguments, configures CMake build flags based on environment variables for different hardware backends (CUDA, SYCL, Vulkan, ROCm, Metal, WebGPU, MUSA, KleidiAI), builds the project, and runs tests. For CUDA builds, it auto-detects GPU architecture via nvidia-smi. Results including logs and exit codes are written to the output directory.

Usage

Used by self-hosted CI runners and for local pre-submission testing by contributors. Set the appropriate GG_BUILD_* environment variables to enable specific hardware backend builds.

Code Reference

Source Location

• Repository: Ggml_org_Llama_cpp
• File: ci/run.sh
• Lines: 1-709

Signature

#!/usr/bin/env bash
# usage: ci/run.sh <output-dir> <mnt-dir>

# Environment variables:
# GG_BUILD_CUDA=1     - Enable CUDA backend
# GG_BUILD_SYCL=1     - Enable SYCL backend
# GG_BUILD_VULKAN=1   - Enable Vulkan backend
# GG_BUILD_ROCM=1     - Enable ROCm/HIP backend
# GG_BUILD_METAL=1    - Enable Metal backend
# GG_BUILD_WEBGPU=1   - Enable WebGPU backend
# GG_BUILD_MUSA=1     - Enable MUSA backend
# GG_BUILD_KLEIDIAI=1 - Enable KleidiAI backend

Import

# CPU-only build
bash ./ci/run.sh ./tmp/results ./tmp/mnt

# With CUDA support
GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

I/O Contract

Inputs

Name	Type	Required	Description
output-dir	path (arg $1)	Yes	Directory where build logs and exit codes are written
mnt-dir	path (arg $2)	Yes	Mount directory for test data and temporary files
GG_BUILD_CUDA	env var	No	Set to 1 to enable CUDA backend build
GG_BUILD_SYCL	env var	No	Set to 1 to enable SYCL backend build
GG_BUILD_VULKAN	env var	No	Set to 1 to enable Vulkan backend build
GG_BUILD_ROCM	env var	No	Set to 1 to enable ROCm/HIP backend build
GG_BUILD_METAL	env var	No	Set to 1 to enable Metal backend build
GG_BUILD_AMDGPU_TARGETS	env var	No	GPU architecture for ROCm builds (e.g., gfx90a, gfx1100)

Outputs

Name	Type	Description
*.log	file	Build and test logs for each test step
*.exit	file	Exit code files indicating pass/fail for each test step
*.md	file	Markdown formatted test result summaries

Usage Examples

# Create output directories
mkdir -p tmp/results tmp/mnt

# Run CPU-only CI build and tests
bash ./ci/run.sh ./tmp/results ./tmp/mnt

# Run with CUDA support (auto-detects GPU architecture)
GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

# Run with Vulkan support
GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

Related Pages

• Principle:Ggml_org_Llama_cpp_CI_Testing