Implementation:Ggml org Ggml Ci runner

Implementation Metadata

File Name	ci/run.sh
Repository	ggml-org/ggml
Lines	395
Language	Shell (Bash)
Domain Tags	CI_CD, Build_System, Testing
Status	Active
Last Updated	2025-05-15 12:00 GMT
Knowledge Sources	ggml-org/ggml repository

Overview

ci/run.sh is the CI build and test runner script supporting multiple GPU backends. It provides a unified CI entry point for building and testing GGML across all supported hardware accelerator backends (CPU, CUDA, Metal, ROCm, SYCL, Vulkan, WebGPU, MUSA), used in both local and automated CI environments.

Description

The script accepts output and mount directories as arguments and detects GPU backend configuration via environment variables:

• GG_BUILD_METAL -- Enables Metal backend (-DGGML_METAL=ON)
• GG_BUILD_CUDA -- Enables CUDA with auto-detected GPU architecture via nvidia-smi
• GG_BUILD_ROCM -- Enables HIP/ROCm with required GG_BUILD_AMDGPU_TARGETS
• GG_BUILD_SYCL -- Enables SYCL with oneAPI environment setup (icx/icpx compilers)
• GG_BUILD_VULKAN -- Enables Vulkan (disables Metal on macOS)
• GG_BUILD_WEBGPU -- Enables WebGPU backend
• GG_BUILD_MUSA -- Enables MUSA backend for Moore Threads GPUs

The workflow:

1. Cleans previous output logs
2. Configures CMake with backend-specific flags
3. Builds with cmake
4. Runs tests via ctest
5. Logs results to the output directory

Usage

mkdir -p tmp/results tmp/mnt

# 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

# With SYCL support
GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

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

Code Reference

Source Location

Repository	File	Lines
ggml-org/ggml	ci/run.sh	395

Key Signatures

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

# Environment variables controlling backend selection:
# GG_BUILD_METAL, GG_BUILD_CUDA, GG_BUILD_ROCM,
# GG_BUILD_SYCL, GG_BUILD_VULKAN, GG_BUILD_WEBGPU, GG_BUILD_MUSA

# CUDA architecture auto-detection:
CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')

# SYCL environment setup:
export ONEAPI_DEVICE_SELECTOR="level_zero:0"
export ZES_ENABLE_SYSMAN=1

I/O Contract

Inputs

• $1 -- Output directory for build logs and results
• $2 -- Mount directory for test data
• Environment variables -- Backend selection flags

Outputs

• Build artifacts -- Compiled GGML library and tests
• Test results -- ctest output logs in output directory
• Exit codes -- Per-test exit status files (*.exit)

Usage Examples

Full CI pipeline:

# 1. Create output directories
mkdir -p results mnt

# 2. Run CI with CUDA
GG_BUILD_CUDA=1 bash ./ci/run.sh ./results ./mnt

# 3. Check results
ls results/*.log    # Build and test logs
ls results/*.exit   # Per-test exit codes
ls results/*.md     # Summary reports

Related Pages

Implements Principle

• Principle:Ggml_org_Ggml_CI_Build_Testing

Related Implementations

• Implementation:Ggml_org_Ggml_Ggml_backend_load_all -- Backend loading tested by CI