Skip to content

Add AMD ROCm GPU build and test CI infrastructure #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Geramy merged 16 commits into from
Apr 17, 2026
Merged

Add AMD ROCm GPU build and test CI infrastructure #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
c8a540d
Add AMD ROCm GPU build and test CI infrastructure
Geramy Mar 3, 2026
91fa40d
I have removed duplicated steps, stray echo, all dead code and utputs…
Geramy Mar 4, 2026
265c868
Replaced grep -oP '(?<=<Key>)[^<]*' with portable sed -n 's/.*<Key>\(…
Geramy Mar 4, 2026
2c853b7
Reverted the change using cmake instead of make adding mode.
Geramy Mar 4, 2026
2fd0fb7
make compatible with git bash
Geramy Mar 9, 2026
4d25bcf
changing the approach to use awk instead, since prior attempt with tr…
Geramy Mar 9, 2026
f1287e3
attempting grep now.
Geramy Mar 9, 2026
cd3b5fc
adding powershell job for winodws instead of shell because shells gre…
Geramy Mar 9, 2026
73de19c
fix: address PR review comments from ramkrishna2910
Geramy Apr 16, 2026
93276f0
refactor: consolidate ROCm version resolution to shared bash script
Geramy Apr 16, 2026
9a4b110
fix(ci): fix Windows ROCm latest auto-detection regex compatibility
Geramy Apr 17, 2026
fbe51d9
Fix echo command for ROCm URL output
Geramy Apr 17, 2026
e071905
update version_regex to be windows and linux compatbile.
Geramy Apr 17, 2026
62589cb
refactor(ci): replace Amazon S3 nightly scanning with AMD official repo
Geramy Apr 17, 2026
8ce7027
fix(ci): allow custom ROCm version input in workflow_dispatch
Geramy Apr 17, 2026
cf3281c
Merge branch 'master' of https://github.com/lemonade-sdk/whisper.cpp-…
Copilot Apr 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
248 changes: 33 additions & 215 deletions .github/workflows/build.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,9 +50,9 @@ on:
required: false
default: 'gfx1151,gfx1150,gfx120X,gfx110X'
rocm_version:
description: 'ROCm version to use (e.g., 7.11.0a20251205) or "latest" to auto-detect'
description: 'ROCm version to use (e.g., 7.12.0, 7.2.1). Available versions: https://repo.amd.com/rocm/tarball/'
required: false
default: 'latest'
default: '7.12.0'

concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
Expand All @@ -66,7 +66,7 @@ env:
ubuntu_image: "ubuntu:22.04"
VCPKG_BINARY_SOURCES: "clear;x-gha,readwrite"
GFX_TARGETS: ${{ github.event.inputs.gfx_target || 'gfx1151,gfx1150,gfx120X,gfx110X' }}
ROCM_VERSION: ${{ github.event.inputs.rocm_version || 'latest' }}
ROCM_VERSION: ${{ github.event.inputs.rocm_version || '7.12.0' }}

jobs:
determine-tag:
Expand Down Expand Up @@ -127,17 +127,12 @@ jobs:
outputs:
windows_matrix: ${{ steps.set-matrix.outputs.windows_matrix }}
ubuntu_matrix: ${{ steps.set-matrix.outputs.ubuntu_matrix }}
should_build_windows: ${{ steps.set-matrix.outputs.should_build_windows }}
should_build_ubuntu: ${{ steps.set-matrix.outputs.should_build_ubuntu }}
steps:
- name: Set matrix
id: set-matrix
run: |
targets="${{ env.GFX_TARGETS }}"
operating_systems="windows,ubuntu"

echo "Input targets: $targets"
echo "Input operating systems: $operating_systems"

target_array=$(echo "$targets" \
| tr ',' '\n' \
Expand All @@ -149,31 +144,11 @@ jobs:
windows_matrix=$(echo "$target_array" \
| jq -c '{gfx_target: ., sdl2: ["ON"], build: ["Release"], arch: ["x64"], s2arc: ["x64"], s2ver: ["2.28.5"]}')

# 3. Create Ubuntu Matrix
ubuntu_matrix=$(echo "$target_array" \
| jq -c '{gfx_target: ., sdl2: ["ON"], build: ["Release"], arch: ["linux/amd64"]}')

# Check which operating systems to build
should_build_windows="false"
should_build_ubuntu="false"

if [[ "$operating_systems" == *"windows"* ]]; then
should_build_windows="true"
echo "windows_matrix=$windows_matrix" >> $GITHUB_OUTPUT
fi

if [[ "$operating_systems" == *"ubuntu"* ]]; then
should_build_ubuntu="true"
echo "ubuntu_matrix=$ubuntu_matrix" >> $GITHUB_OUTPUT
fi

echo "should_build_windows=$should_build_windows" >> $GITHUB_OUTPUT
echo "should_build_ubuntu=$should_build_ubuntu" >> $GITHUB_OUTPUT

echo "Windows build: $should_build_windows"
echo "Ubuntu build: $should_build_ubuntu"
echo "Generated matrix: $matrix_targets"
#linux/amd64
echo "windows_matrix=$windows_matrix" >> $GITHUB_OUTPUT
echo "ubuntu_matrix=$ubuntu_matrix" >> $GITHUB_OUTPUT

ubuntu-22:
if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
Expand Down Expand Up @@ -469,17 +444,20 @@ jobs:
github.event.inputs.run_type == 'full-ci' }}
runs-on: ubuntu-22.04
needs: [determine-tag, prepare-matrix]
# Check if we should run (based on the prepare-matrix output)
strategy:
# Uses the matrix generated in prepare-matrix (gfx_target, arch, build, sdl2)
matrix: ${{fromJson(needs.prepare-matrix.outputs.ubuntu_matrix)}}
fail-fast: false
outputs:
rocm_version: ${{ steps.set-outputs.outputs.rocm_version }}

steps:
- name: Free disk space
run: curl -fsSL https://raw.githubusercontent.com/kou/arrow/e49d8ae15583ceff03237571569099a6ad62be32/ci/scripts/util_free_space.sh | bash
run: |
echo "=== Disk usage before cleanup ==="
df -h /
sudo rm -rf /usr/local/lib/android /opt/ghc /usr/local/share/boost \
/usr/share/dotnet /usr/local/.ghcup /opt/hostedtoolcache/CodeQL
sudo docker image prune --all --force 2>/dev/null || true
echo "=== Disk usage after cleanup ==="
df -h /

- name: Checkout repository
uses: actions/checkout@v4
Expand All @@ -491,86 +469,10 @@ jobs:

- name: Download and extract ROCm directly to /opt/rocm
run: |
# Determine ROCm version to use
rocm_version="${{ env.ROCM_VERSION }}"
current_target="${{ matrix.gfx_target }}"

# Add appropriate suffixes for different GPU targets
s3_target="$current_target"
if [ "$current_target" = "gfx110X" ]; then
s3_target="${current_target}-dgpu"
echo "Using S3 target with -dgpu suffix: $s3_target"
elif [ "$current_target" = "gfx120X" ]; then
s3_target="${current_target}-all"
echo "Using S3 target with -all suffix: $s3_target"
fi

if [ "$rocm_version" = "latest" ]; then
echo "Auto-detecting latest ROCm version for target: $current_target"
s3_response=$(curl -s "https://therock-nightly-tarball.s3.amazonaws.com/?prefix=therock-dist-linux-${s3_target}-7")

# Extract all files
files=$(echo "$s3_response" | grep -oP '(?<=<Key>)[^<]*' | grep "therock-dist-linux-${s3_target}-")

# Extract versions and sort them properly
latest_file=""
latest_major=0
latest_minor=0
latest_patch=0
latest_rc=0
latest_is_alpha=false

while IFS= read -r file; do
if [[ "$file" =~ therock-dist-linux-${s3_target}-.*?([0-9]+\.[0-9]+\.[0-9]+(a|rc)[0-9]+)\.tar\.gz ]]; then
version="${BASH_REMATCH[1]}"
major=$(echo "$version" | cut -d. -f1)
minor=$(echo "$version" | cut -d. -f2)
patch=$(echo "$version" | cut -d. -f3 | sed 's/\(a\|rc\).*//')
rc=$(echo "$version" | sed 's/.*\(a\|rc\)//')
is_alpha=false
if [[ "$version" =~ a ]]; then is_alpha=true; fi

is_newer=false
if [ "$major" -gt "$latest_major" ]; then is_newer=true;
elif [ "$major" -eq "$latest_major" ] && [ "$minor" -gt "$latest_minor" ]; then is_newer=true;
elif [ "$major" -eq "$latest_major" ] && [ "$minor" -eq "$latest_minor" ] && [ "$patch" -gt "$latest_patch" ]; then is_newer=true;
elif [ "$major" -eq "$latest_major" ] && [ "$minor" -eq "$latest_minor" ] && [ "$patch" -eq "$latest_patch" ]; then
if [ "$is_alpha" = true ] && [ "$latest_is_alpha" = false ]; then is_newer=true;
elif [ "$is_alpha" = "$latest_is_alpha" ] && [ "$rc" -gt "$latest_rc" ]; then is_newer=true;
fi
fi

if [ "$is_newer" = true ]; then
latest_file="$file"
latest_major="$major"
latest_minor="$minor"
latest_patch="$patch"
latest_rc="$rc"
latest_is_alpha="$is_alpha"
fi
fi
done <<< "$files"

echo "Found latest file: $latest_file"

if [[ "$latest_file" =~ therock-dist-linux-${s3_target}-.*?([0-9]+\.[0-9]+\.[0-9]+(a|rc)[0-9]+)\.tar\.gz ]]; then
rocm_version="${BASH_REMATCH[1]}"
echo "Detected latest ROCm version: $rocm_version"
else
echo "Failed to extract ROCm version from latest file: $latest_file"
exit 1
fi

rocm_url="https://therock-nightly-tarball.s3.amazonaws.com/$latest_file"
else
rocm_url="https://therock-nightly-tarball.s3.amazonaws.com/therock-dist-linux-${s3_target}-${rocm_version}.tar.gz"
fi

echo "DETECTED_ROCM_VERSION=$rocm_version" >> $GITHUB_ENV

# Create directory and stream extraction
source ci/resolve-rocm-version.sh linux "${{ matrix.gfx_target }}" "${{ env.ROCM_VERSION }}"
echo "DETECTED_ROCM_VERSION=$ROCM_RESOLVED_VERSION" >> $GITHUB_ENV
sudo mkdir -p /opt/rocm
curl -sL "$rocm_url" | sudo tar --use-compress-program=gzip -xf - -C /opt/rocm --strip-components=1
curl -sL "$ROCM_TARBALL_URL" | sudo tar --use-compress-program=gzip -xf - -C /opt/rocm --strip-components=1

- name: Set ROCm environment variables
run: |
Expand All @@ -595,34 +497,17 @@ jobs:

- name: Configure CMake
run: |

# Map GPU targets
current_target="${{ matrix.gfx_target }}"
echo "Input target: $current_target"

if [ "$current_target" = "gfx110X" ]; then
mapped_target="gfx1100;gfx1101;gfx1102"
elif [ "$current_target" = "gfx1151" ]; then
mapped_target="gfx1151"
elif [ "$current_target" = "gfx1150" ]; then
mapped_target="gfx1150"
elif [ "$current_target" = "gfx120X" ]; then
mapped_target="gfx1200;gfx1201"
else
mapped_target="$current_target"
fi
echo "Mapped target: $mapped_target"
source ci/map-gpu-target.sh "${{ matrix.gfx_target }}"

cmake -S . -B build -G Ninja \
-DCMAKE_C_COMPILER=/opt/rocm/llvm/bin/clang \
-DCMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ \
-DCMAKE_HIP_FLAGS="--rocm-path=/opt/rocm --rocm-device-lib-path=${{ env.ROCM_BITCODE_PATH }}" \
-DCMAKE_PREFIX_PATH=/opt/rocm \
-DCMAKE_BUILD_TYPE=${{ matrix.build }} \
-DGPU_TARGETS="$mapped_target" \
-DGPU_TARGETS="$MAPPED_GPU_TARGET" \
-DBUILD_SHARED_LIBS=ON \
-DGGML_HIP=ON \
-DGGML_ROCM=1 \
-DWHISPER_SDL2=${{ matrix.sdl2 }}

- name: Build
Expand Down Expand Up @@ -724,7 +609,7 @@ jobs:
zip -r ../../whisper-bin-${{ matrix.gfx_target }}-${SAFE_ARCH}.zip .

- name: Upload binaries
if: matrix.sdl2 == 'ON' && needs.determine-tag.outputs.should_release
if: ${{ matrix.sdl2 == 'ON' && needs.determine-tag.outputs.should_release }}
uses: actions/upload-artifact@v4
with:
name: whisper-bin-${{ matrix.gfx_target }}-${{ env.SAFE_ARCH }}
Expand Down Expand Up @@ -1313,8 +1198,6 @@ jobs:
strategy:
matrix: ${{fromJson(needs.prepare-matrix.outputs.windows_matrix)}}
fail-fast: false
outputs:
rocm_version: ${{ steps.set-outputs.outputs.rocm_version }}

steps:
- name: Clone
Expand Down Expand Up @@ -1371,96 +1254,37 @@ jobs:
exit 1
}

- name: Download ROCm nightly tarball
- name: Resolve ROCm version and download tarball
Comment thread
Geramy marked this conversation as resolved.
shell: bash
run: |
# Determine ROCm version to use
$rocmVersion = "${{ env.ROCM_VERSION }}"
$currentTarget = "${{ matrix.gfx_target }}"

$s3Target = $currentTarget
if ($currentTarget -eq "gfx110X") {
$s3Target = "$currentTarget-dgpu"
Write-Host "Using S3 target with -dgpu suffix: $s3Target"
} elseif ($currentTarget -eq "gfx120X") {
$s3Target = "$currentTarget-all"
Write-Host "Using S3 target with -all suffix: $s3Target"
}

if ($rocmVersion -eq "latest") {
Write-Host "Auto-detecting latest ROCm version for target: $currentTarget"
$s3Response = (Invoke-WebRequest "https://therock-nightly-tarball.s3.amazonaws.com/?prefix=therock-dist-windows-$s3Target-7").Content
$files = $s3Response -split '<Key>' | Where-Object {$_ -match '</Key>'} | ForEach-Object { ($_ -split '</Key>')[0] }

$versionFiles = @()
foreach ($file in $files) {
if ($file -match "therock-dist-windows-$s3Target-.*?(\d+\.\d+\.\d+(?:a|rc)\d+)\.tar\.gz") {
$version = $matches[1]
$versionFiles += [PSCustomObject]@{
File = $file
Version = $version
Major = [int]($version -split '\.')[0]
Minor = [int]($version -split '\.')[1]
Patch = [int](($version -split '\.')[2] -replace '(?:a|rc).*', '')
RC = [int]($version -replace '.*(?:a|rc)', '')
IsAlpha = $version -match 'a'
}
}
}

$latestFile = ($versionFiles | Sort-Object Major, Minor, Patch, @{Expression={if($_.IsAlpha){1}else{0}}}, RC | Select-Object -Last 1).File
Write-Host "Found latest file: $latestFile"

if ($latestFile -match "therock-dist-windows-$s3Target-.*?(\d+\.\d+\.\d+(?:a|rc)\d+)\.tar\.gz") {
$rocmVersion = $matches[1]
Write-Host "Detected latest ROCm version: $rocmVersion"
} else {
Write-Error "Failed to extract ROCm version from latest file: $latestFile"
exit 1
}
$rocmUrl = "https://therock-nightly-tarball.s3.amazonaws.com/$latestFile"
} else {
$rocmUrl = "https://therock-nightly-tarball.s3.amazonaws.com/therock-dist-windows-$s3Target-$rocmVersion.tar.gz"
}

echo "DETECTED_ROCM_VERSION=$rocmVersion" >> $env:GITHUB_ENV
Invoke-WebRequest -Uri $rocmUrl -OutFile "rocm.tar.gz"
source ci/resolve-rocm-version.sh windows "${{ matrix.gfx_target }}" "${{ env.ROCM_VERSION }}"
echo "DETECTED_ROCM_VERSION=$ROCM_RESOLVED_VERSION" >> $GITHUB_ENV
curl -sL "$ROCM_TARBALL_URL" -o rocm.tar.gz

- name: Extract ROCm to C:\opt\rocm
run: |
New-Item -ItemType Directory -Force -Path "C:\opt\rocm"
tar -xzf rocm.tar.gz -C C:\opt\rocm --strip-components=1

- name: Map GPU target
id: gpu-target
shell: bash
run: |
source ci/map-gpu-target.sh "${{ matrix.gfx_target }}"
echo "mapped=$MAPPED_GPU_TARGET" >> $GITHUB_OUTPUT

- name: Configure
shell: pwsh
run: |
$currentTarget = "${{ matrix.gfx_target }}"
Write-Host "Input target: $currentTarget"

if ($currentTarget -eq "gfx110X") {
$mapped_target = "gfx1100;gfx1101;gfx1102"
} elseif ($currentTarget -eq "gfx1151") {
$mapped_target = "gfx1151"
} elseif ($currentTarget -eq "gfx1150") {
$mapped_target = "gfx1150"
} elseif ($currentTarget -eq "gfx120X") {
$mapped_target = "gfx1200;gfx1201"
} else {
$mapped_target = $currentTarget
}
Write-Host "Mapped target: $mapped_target"

# Set up environment variables and PATH
$env:HIP_PATH = "C:\opt\rocm"
$env:HIP_PLATFORM = "amd"
# Ensure bin comes before llvm\bin for consistency
$env:PATH = "$env:HIP_PATH\bin;$env:HIP_PATH\lib\llvm\bin;$env:PATH"

# Define CMake arguments
$cmakeArgs = @(
"-S", ".",
"-B", "build",
"-G", "Ninja Multi-Config",
"-DGPU_TARGETS=$mapped_target",
"-DGPU_TARGETS=${{ steps.gpu-target.outputs.mapped }}",
"-DGGML_HIP=ON",
"-DCMAKE_C_COMPILER=$env:HIP_PATH/lib/llvm/bin/amdclang.exe",
"-DCMAKE_CXX_COMPILER=$env:HIP_PATH/lib/llvm/bin/amdclang++.exe",
Expand All @@ -1469,12 +1293,10 @@ jobs:
"-DCMAKE_CXX_FLAGS='-D__PRFCHWINTRIN_H'",
"-DCMAKE_HIP_FLAGS=--rocm-path=C:/opt/rocm",
"-DCMAKE_PREFIX_PATH=$env:HIP_PATH",
"-DGGML_ROCM=1",
"-DCMAKE_BUILD_TYPE=${{ matrix.build }}",
"-DBUILD_SHARED_LIBS=ON",
"-DWHISPER_SDL2=${{ matrix.sdl2 }}"
)
# Run CMake
cmake @cmakeArgs

- name: Build
Expand Down Expand Up @@ -1528,10 +1350,6 @@ jobs:
}
}

- name: Copy SDL2.dll
if: matrix.sdl2 == 'ON'
run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}

- name: Copy SDL2.dll
if: matrix.sdl2 == 'ON'
run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" "build/bin/${{ matrix.build }}"
Expand All @@ -1544,7 +1362,7 @@ jobs:
Compress-Archive -Path "build/bin/${{ matrix.build }}" -DestinationPath $zipName

- name: Upload binaries
if: matrix.sdl2 == 'ON' && needs.determine-tag.outputs.should_release
if: ${{ matrix.sdl2 == 'ON' && needs.determine-tag.outputs.should_release }}
uses: actions/upload-artifact@v4
with:
# Unique artifact name per matrix job
Expand Down
Loading
Loading