perf(stft): optimize FFT with table-based transforms and memory pooling #44

Workflow file for this run

	name: CI

	on:
	push:
	branches: [main, master]
	paths:
	- '*/.cpp'
	- '*/.h'
	- '*/.hpp'
	- '**/CMakeLists.txt'
	- '.github/workflows/**'
	pull_request:
	types: [opened, synchronize, reopened]
	workflow_dispatch:

	concurrency:
	group: ${{ github.workflow }}-${{ github.head_ref && github.ref \|\| github.run_id }}
	cancel-in-progress: true

	env:
	# HuggingFace model info (Mel-Band)
	HF_MB_REPO: GaboxR67/MelBandRoformers
	HF_MB_CHECKPOINT: melbandroformers/vocals/voc_fv6.ckpt
	HF_MB_CONFIG: melbandroformers/vocals/voc_gabox.yaml
	# BS Roformer model info
	HF_BS_REPO: anvuew/BS-RoFormer
	HF_BS_CHECKPOINT: bs_roformer_anvuew_sdr_12.45.ckpt
	HF_BS_CONFIG: config.yaml
	# Music-Source-Separation-Training repo
	MSST_REPO: https://github.com/ZFTurbo/Music-Source-Separation-Training.git
	# Enable sccache GitHub Actions cache
	SCCACHE_GHA_ENABLED: "true"

	jobs:
	# ===========================================================================
	# Prepare: Generate test data (runs once, shared via artifacts)
	# ===========================================================================
	prepare-test-data:
	runs-on: ubuntu-latest

	steps:
	- name: Checkout
	uses: actions/checkout@v4

	- name: Setup Python
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Clone MSST Repository
	run: git clone --depth 1 ${{ env.MSST_REPO }} msst

	- name: Install Dependencies
	run: \|
	pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu
	pip install huggingface_hub scipy soundfile gguf librosa ml_collections einops pyyaml numpy tqdm beartype rotary_embedding_torch

	- name: Download Model from HuggingFace
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	run: \|
	python -c "
	from huggingface_hub import hf_hub_download
	import os
	token = os.environ.get('HF_TOKEN') or None

	# Download Mel-Band Roformer
	hf_hub_download('${{ env.HF_MB_REPO }}', '${{ env.HF_MB_CHECKPOINT }}',
	local_dir='./model', token=token)
	hf_hub_download('${{ env.HF_MB_REPO }}', '${{ env.HF_MB_CONFIG }}',
	local_dir='./model', token=token)
	# Download BS Roformer
	hf_hub_download('${{ env.HF_BS_REPO }}', '${{ env.HF_BS_CHECKPOINT }}',
	local_dir='./model_bs', token=token)
	hf_hub_download('${{ env.HF_BS_REPO }}', '${{ env.HF_BS_CONFIG }}',
	local_dir='./model_bs', token=token)
	"

	- name: Generate Test Audio
	run: \|
	python scripts/generate_test_audio.py --output test_audio.wav --duration 5.0 --sample-rate 44100

	- name: Generate Test Data
	run: \|
	python scripts/generate_test_data.py \
	--model-repo msst \
	--audio test_audio.wav \
	--checkpoint model/${{ env.HF_MB_CHECKPOINT }} \
	--config model/${{ env.HF_MB_CONFIG }} \
	--output test_data

	- name: Convert Model to GGUF (Mel-Band)
	run: \|
	python scripts/convert_to_gguf.py \
	--ckpt model/${{ env.HF_MB_CHECKPOINT }} \
	--config model/${{ env.HF_MB_CONFIG }} \
	--out model.gguf \
	--dtype fp32

	# ----- BS Roformer Setup -----
	- name: Generate Test Data (BS)
	run: \|
	# Need to make sure msst is in python path
	export PYTHONPATH=$PYTHONPATH:$(pwd)/msst
	# Use real BS model
	python scripts/generate_test_data.py \
	--model-repo msst \
	--audio test_audio.wav \
	--checkpoint model_bs/${{ env.HF_BS_CHECKPOINT }} \
	--config model_bs/${{ env.HF_BS_CONFIG }} \
	--output test_data_bs

	- name: Convert Model to GGUF (BS)
	run: \|
	python scripts/convert_to_gguf.py \
	--ckpt model_bs/${{ env.HF_BS_CHECKPOINT }} \
	--config model_bs/${{ env.HF_BS_CONFIG }} \
	--out model_bs.gguf \
	--dtype fp32 \
	--arch bs

	- name: Upload Test Data Artifact
	uses: actions/upload-artifact@v4
	with:
	name: test-data
	path: \|
	test_data/
	model.gguf
	test_audio.wav
	test_data_bs/
	model_bs.gguf
	retention-days: 1

	# ===========================================================================
	# Build Matrix: Core Platforms + Vulkan
	# ===========================================================================
	build:
	needs: prepare-test-data
	strategy:
	fail-fast: false
	matrix:
	include:
	# Tier 1: Core Platforms (CPU)
	- { name: linux-x64-cpu, os: ubuntu-22.04, backend: cpu, test: true }
	- { name: linux-arm64-cpu, os: ubuntu-22.04-arm, backend: cpu, test: true }
	- { name: macos-arm64, os: macos-latest, backend: cpu, test: true }
	- { name: macos-x64, os: macos-15-intel, backend: cpu, test: true }
	- { name: windows-x64-msvc, os: windows-2025, backend: cpu, test: true }
	# Tier 2: Vulkan Backend
	- { name: linux-vulkan, os: ubuntu-24.04, backend: vulkan, test: true }
	- { name: windows-vulkan, os: windows-2025, backend: vulkan, test: true }

	runs-on: ${{ matrix.os }}

	steps:
	- name: Checkout
	uses: actions/checkout@v4

	- name: Clone GGML
	run: git clone --depth 1 https://github.com/ggerganov/ggml.git ggml

	- name: Download Test Data
	uses: actions/download-artifact@v4
	with:
	name: test-data

	- name: Setup Python
	uses: actions/setup-python@v5
	with:
	python-version: '3.11'

	- name: Setup MSVC
	if: runner.os == 'Windows'
	uses: ilammy/msvc-dev-cmd@v1

	- name: Install Python Dependencies
	run: pip install numpy scipy

	- name: Setup sccache
	uses: mozilla-actions/sccache-action@v0.0.9
	with:
	version: "v0.12.0"

	# ----- Linux Dependencies -----
	- name: Install Dependencies (Linux)
	if: runner.os == 'Linux'
	run: \|
	sudo apt-get update
	sudo apt-get install -y build-essential cmake

	- name: Install Vulkan SDK (Linux)
	if: matrix.backend == 'vulkan' && runner.os == 'Linux'
	run: \|
	sudo apt-get install -y libvulkan-dev glslc mesa-vulkan-drivers

	# ----- macOS Dependencies -----
	- name: Install Dependencies (macOS)
	if: runner.os == 'macOS'
	run: brew install cmake

	# ----- Windows Dependencies -----
	- name: Install Dependencies (Windows)
	if: runner.os == 'Windows'
	run: choco install ninja -y

	- name: Install Vulkan SDK (Windows)
	if: matrix.backend == 'vulkan' && runner.os == 'Windows'
	run: \|
	$VK_VERSION = "1.4.313.2"
	curl.exe -o VulkanSDK.exe -L "https://sdk.lunarg.com/sdk/download/${VK_VERSION}/windows/vulkansdk-windows-X64-${VK_VERSION}.exe"
	Start-Process -FilePath .\VulkanSDK.exe -ArgumentList "--accept-licenses --default-answer --confirm-command install" -Wait
	Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${VK_VERSION}"
	Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${VK_VERSION}\bin"

	# ----- Configure -----
	- name: Configure (Unix)
	if: runner.os != 'Windows'
	run: \|
	cmake -B build \
	-DCMAKE_BUILD_TYPE=Release \
	-DGGML_DIR=ggml \
	-DGGML_CUDA=OFF \
	-DGGML_VULKAN=${{ matrix.backend == 'vulkan' && 'ON' \|\| 'OFF' }} \
	-DCMAKE_C_COMPILER_LAUNCHER=sccache \
	-DCMAKE_CXX_COMPILER_LAUNCHER=sccache \
	-DMBR_BUILD_TESTS=ON \
	-DMBR_BUILD_CLI=ON

	- name: Configure (Windows)
	if: runner.os == 'Windows'
	run: \|
	cmake -B build -G "Ninja Multi-Config" `
	-DGGML_DIR=ggml `
	-DGGML_CUDA=OFF `
	-DGGML_VULKAN=${{ matrix.backend == 'vulkan' && 'ON' \|\| 'OFF' }} `
	-DCMAKE_C_COMPILER_LAUNCHER=sccache `
	-DCMAKE_CXX_COMPILER_LAUNCHER=sccache `
	-DMBR_BUILD_TESTS=ON `
	-DMBR_BUILD_CLI=ON

	# ----- Build -----
	- name: Build (Unix)
	if: runner.os != 'Windows'
	run: cmake --build build --config Release -j $(nproc 2>/dev/null \|\| sysctl -n hw.logicalcpu)

	- name: Build (Windows)
	if: runner.os == 'Windows'
	run: cmake --build build --config Release -j $env:NUMBER_OF_PROCESSORS

	# ----- Unit Tests -----
	- name: Run Unit Tests
	if: matrix.test
	env:
	MBR_MODEL_PATH: ${{ github.workspace }}/model.gguf
	MBR_TEST_DATA_DIR: ${{ github.workspace }}/test_data
	MBR_FORCE_CPU: ${{ runner.os == 'macOS' && '1' \|\| '' }}
	run: ctest --test-dir build -C Release -V --output-on-failure --timeout 300

	- name: Run Unit Tests (BS Roformer)
	if: matrix.test
	env:
	MBR_MODEL_PATH: ${{ github.workspace }}/model_bs.gguf
	MBR_TEST_DATA_DIR: ${{ github.workspace }}/test_data_bs
	MBR_ARCHITECTURE: bs
	MBR_FORCE_CPU: ${{ runner.os == 'macOS' && '1' \|\| '' }}
	run: ctest --test-dir build -C Release -V --output-on-failure --timeout 300

	# ----- CLI Tests -----
	- name: Test CLI
	if: matrix.test
	shell: bash
	env:
	MBR_MODEL_PATH: ${{ github.workspace }}/model.gguf
	MBR_FORCE_CPU: ${{ runner.os == 'macOS' && '1' \|\| '' }}
	run: \|
	echo "=== CLI Test Suite ==="

	# Dynamically find CLI executable
	CLI_NAME="bs_roformer-cli"
	if [[ "$RUNNER_OS" == "Windows" ]]; then CLI_NAME="bs_roformer-cli.exe"; fi

	echo "Searching for $CLI_NAME..."
	CLI_PATH=$(find build -name "$CLI_NAME" \| head -n 1)

	if [[ -z "$CLI_PATH" ]]; then
	echo "Error: CLI executable not found!"
	find build
	exit 1
	fi

	echo "Found CLI at: $CLI_PATH"
	chmod +x "$CLI_PATH"

	# Setup execution variables
	CLI_DIR=$(dirname "$CLI_PATH")
	CLI_EXE="./$CLI_NAME"

	# Debug Info
	echo "Debug: Listing directory $CLI_DIR"
	ls -l "$CLI_DIR"

	echo "Debug: Checking dependencies"
	if command -v ldd >/dev/null; then
	ldd "$CLI_PATH" \|\| echo "ldd returned error"
	fi

	# Define runner helper
	run_cli() {
	echo "Executing: ./$CLI_NAME $@"
	(cd "$CLI_DIR" && ./$CLI_NAME "$@")
	}

	# Debug PATH
	echo "PATH: $PATH"

	# 1. Test --help
	echo "[1/4] Testing --help..."
	run_cli --help

	# 2. Test with missing arguments (should fail)
	echo "[2/4] Testing error handling..."
	if run_cli 2>/dev/null; then
	echo "ERROR: CLI should fail without arguments"
	exit 1
	fi

	# 3. Use existing test audio
	echo "[3/4] Using generated test audio..."
	cp test_audio.wav cli_test_input.wav

	# 4. Run full inference
	echo "[4/4] Running inference..."
	# Use absolute paths for input/output to avoid directory issues
	ABS_MODEL=$(readlink -f "$MBR_MODEL_PATH" \|\| echo "$(pwd)/model.gguf")
	ABS_INPUT=$(readlink -f cli_test_input.wav \|\| echo "$(pwd)/cli_test_input.wav")
	ABS_OUTPUT=$(readlink -f cli_test_output.wav \|\| echo "$(pwd)/cli_test_output.wav")

	# Convert paths for Windows MSVC executables
	if [[ "$RUNNER_OS" == "Windows" ]]; then
	ABS_MODEL=$(cygpath -w "$ABS_MODEL")
	ABS_INPUT=$(cygpath -w "$ABS_INPUT")
	ABS_OUTPUT=$(cygpath -w "$ABS_OUTPUT")
	fi

	echo "Running with model: $ABS_MODEL"
	run_cli "$ABS_MODEL" "$ABS_INPUT" "$ABS_OUTPUT" --chunk-size 88200 --overlap 2

	# Verify output exists and has reasonable size
	if [[ ! -f cli_test_output.wav ]]; then
	echo "ERROR: Output file not created"
	exit 1
	fi

	OUTPUT_SIZE=$(stat -c%s cli_test_output.wav 2>/dev/null \|\| stat -f%z cli_test_output.wav)
	if [[ $OUTPUT_SIZE -lt 1000 ]]; then
	echo "ERROR: Output file too small: $OUTPUT_SIZE bytes"
	exit 1
	fi

	echo "=== CLI Tests Passed ==="

	# ----- Upload Artifacts -----
	- name: Upload Build Artifacts
	uses: actions/upload-artifact@v4
	with:
	name: build-${{ matrix.name }}
	path: \|
	build/bin/
	build/lib*/
	build/*.dll
	build/*.so
	build/*.dylib
	build/bs_roformer-cli*
	build/Release/
	retention-days: 7

	# ----- Prepare Release Artifact -----
	- name: Prepare Release Artifact (Unix)
	if: runner.os != 'Windows'
	shell: bash
	run: \|
	# Create release directory
	mkdir -p release/mel-band-roformer

	# Find and copy CLI executable
	CLI_PATH=$(find build -name "bs_roformer-cli" -type f \| head -n 1)
	if [[ -n "$CLI_PATH" ]]; then
	cp "$CLI_PATH" release/mel-band-roformer/
	chmod +x release/mel-band-roformer/bs_roformer-cli
	fi

	# Copy shared libraries if exist (only real files, not symlinks)
	find build $ -name ".so" -o -name "*.dylib" $ -type f ! -type l \| while read lib; do
	cp "$lib" release/mel-band-roformer/ 2>/dev/null \|\| true
	done

	# List contents
	echo "Release artifact contents:"
	ls -lh release/mel-band-roformer/

	- name: Prepare Release Artifact (Windows)
	if: runner.os == 'Windows'
	shell: pwsh
	run: \|
	# Create release directory
	New-Item -ItemType Directory -Force -Path "release\mel-band-roformer"

	# Find and copy CLI executable
	$CliPath = Get-ChildItem -Path build -Filter "bs_roformer-cli.exe" -Recurse -File \| Select-Object -First 1
	if ($CliPath) {
	Copy-Item $CliPath.FullName "release\mel-band-roformer\"
	}

	# Copy DLL files
	Get-ChildItem -Path build -Filter "*.dll" -Recurse -File \| ForEach-Object {
	Copy-Item $_.FullName "release\mel-band-roformer\" -ErrorAction SilentlyContinue
	}

	# List contents
	Write-Host "Release artifact contents:"
	Get-ChildItem "release\mel-band-roformer" \| Format-Table Name, Length

	- name: Upload Release Artifact
	uses: actions/upload-artifact@v4
	with:
	name: BSRoformer-${{ matrix.name }}
	path: release/mel-band-roformer/
	retention-days: 30

	# ===========================================================================
	# CUDA Build: Linux (Compile Verification Only)
	# ===========================================================================
	build-cuda-linux:
	name: build-cuda-linux-${{ matrix.cuda_version }}
	runs-on: ${{ matrix.os }}
	strategy:
	fail-fast: false
	matrix:
	include:
	- { cuda_version: "11.8.0", os: ubuntu-22.04 }
	- { cuda_version: "12.9.1", os: ubuntu-latest }
	- { cuda_version: "13.1.0", os: ubuntu-latest }

	steps:
	- name: Checkout
	uses: actions/checkout@v4

	- name: Clone GGML
	run: git clone --depth 1 https://github.com/ggerganov/ggml.git ggml

	- name: Install CUDA Toolkit
	uses: Jimver/cuda-toolkit@master
	with:
	cuda: ${{ matrix.cuda_version }}
	method: network
	sub-packages: '["nvcc", "cudart", "thrust"]'
	non-cuda-sub-packages: '["libcublas", "libcublas-dev"]'

	- name: Install Dependencies
	run: \|
	sudo apt-get install -y cmake build-essential ninja-build

	- name: Setup sccache
	uses: mozilla-actions/sccache-action@v0.0.9

	- name: Configure
	run: \|
	ls -ld ggml
	# Minimal architectures for compile verification
	# Just verify compilation works, not for distribution
	CUDA_ARCHS="75;86"
	echo "Verifying build for CUDA architectures: $CUDA_ARCHS"
	cmake -B build -G Ninja \
	-DCMAKE_BUILD_TYPE=Release \
	-DGGML_DIR=ggml \
	-DGGML_CUDA=ON \
	-DGGML_CUDA_FORCE_MMQ=ON \
	-DCMAKE_CUDA_ARCHITECTURES="$CUDA_ARCHS" \
	-DCMAKE_C_COMPILER_LAUNCHER=sccache \
	-DCMAKE_CXX_COMPILER_LAUNCHER=sccache \
	-DCMAKE_CUDA_COMPILER_LAUNCHER=sccache \
	-DMBR_BUILD_TESTS=OFF \
	-DMBR_BUILD_CLI=ON

	- name: Build
	run: cmake --build build --config Release -j $(nproc)

	- name: Upload Artifacts
	uses: actions/upload-artifact@v4
	with:
	name: build-linux-cuda-${{ matrix.cuda_version }}
	path: \|
	build/bin/
	build/lib*/
	build/*.so
	retention-days: 7

	# ----- Prepare Release Artifact -----
	- name: Prepare Release Artifact
	run: \|
	# Create release directory
	mkdir -p release/mel-band-roformer

	# Find and copy CLI executable
	CLI_PATH=$(find build -name "bs_roformer-cli" -type f \| head -n 1)
	if [[ -n "$CLI_PATH" ]]; then
	cp "$CLI_PATH" release/mel-band-roformer/
	chmod +x release/mel-band-roformer/bs_roformer-cli
	fi

	# Copy shared libraries
	find build -name ".so" \| while read lib; do
	cp "$lib" release/mel-band-roformer/ 2>/dev/null \|\| true
	done

	# List contents
	echo "Release artifact contents:"
	ls -lh release/mel-band-roformer/

	- name: Upload Release Artifact
	uses: actions/upload-artifact@v4
	with:
	name: BSRoformer-linux-cuda-${{ matrix.cuda_version }}
	path: release/mel-band-roformer/
	retention-days: 30

	# ===========================================================================
	# CUDA Build: Windows (Compile Only - No GPU for testing)
	# ===========================================================================
	build-cuda-windows:
	name: build-cuda-windows-${{ matrix.cuda_version }}
	runs-on: windows-2022
	strategy:
	fail-fast: false
	matrix:
	cuda_version: ["11.8.0", "12.9.1", "13.1.0"]

	env:
	CUDA_VERSION: ${{ matrix.cuda_version }}

	steps:
	- name: Checkout
	uses: actions/checkout@v4

	- name: Setup MSVC
	if: runner.os == 'Windows'
	uses: ilammy/msvc-dev-cmd@v1

	- name: Clone GGML
	run: git clone --depth 1 https://github.com/ggerganov/ggml.git ggml

	- name: Install CUDA Toolkit
	if: ${{ matrix.cuda_version != '13.1.0' }}
	uses: Jimver/cuda-toolkit@master
	with:
	cuda: ${{ matrix.cuda_version }}
	method: network
	sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'

	- name: Install CUDA Toolkit(13.1.0)
	if: ${{ matrix.cuda_version == '13.1.0' }}
	uses: Jimver/cuda-toolkit@master
	with:
	cuda: ${{ matrix.cuda_version }}
	method: network
	sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "nvrtc", "nvrtc_dev", "crt", "nvvm", "visual_studio_integration"]'

	- name: Install Ninja
	run: choco install ninja -y

	- name: Setup sccache
	uses: mozilla-actions/sccache-action@v0.0.9

	- name: Configure and Build
	shell: pwsh
	run: \|
	# Consumer GPU architectures:
	# 61=Pascal (GTX 10), 75=Turing (RTX 20/GTX 16), 86=Ampere (RTX 30), 89=Ada (RTX 40), 120=Blackwell (RTX 50)
	$cudaVersion = "${{ matrix.cuda_version }}"
	if ($cudaVersion -match "^11\.") {
	# CUDA 11.x: Support Pascal (61) to Ada (89)
	$cudaArchs = "61;75;86;89"
	$env:CUDAFLAGS = "-allow-unsupported-compiler -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH -D_DISABLE_CONSTEXPR_MUTEX_CONSTRUCTOR"
	} elseif ($cudaVersion -match "^12\.") {
	# CUDA 12.x: Support Pascal (61) to Blackwell (120) (technically 120 needs 12.8+, keeping unified for now or simple)
	# Actually 12.x still supports Pascal.
	$cudaArchs = "61;75;86;89"
	$env:CUDAFLAGS = ""
	} else {
	# CUDA 13+: Drop Pascal (61). Support Turing (75) to Blackwell (120)
	$cudaArchs = "75;86;89;120"
	$env:CUDAFLAGS = ""
	}
	Write-Host "Building for CUDA architectures: $cudaArchs"

	cmake -B build -G "Ninja Multi-Config" `
	-DGGML_DIR=ggml `
	-DGGML_CUDA=ON `
	-DGGML_CUDA_FORCE_MMQ=ON `
	"-DCMAKE_CUDA_ARCHITECTURES=$cudaArchs" `
	-DCMAKE_C_COMPILER_LAUNCHER=sccache `
	-DCMAKE_CXX_COMPILER_LAUNCHER=sccache `
	-DMBR_BUILD_TESTS=OFF `
	-DMBR_BUILD_CLI=ON
	cmake --build build --config Release -j $env:NUMBER_OF_PROCESSORS

	- name: Upload Artifacts
	uses: actions/upload-artifact@v4
	with:
	name: build-windows-cuda-${{ matrix.cuda_version }}
	path: \|
	build/bin/
	build/Release/
	build/*.dll
	retention-days: 7

	# ----- Prepare Release Artifact -----
	- name: Prepare Release Artifact
	shell: pwsh
	run: \|
	# Create release directory
	New-Item -ItemType Directory -Force -Path "release\mel-band-roformer"

	# Find and copy CLI executable
	$CliPath = Get-ChildItem -Path build -Filter "bs_roformer-cli.exe" -Recurse -File \| Select-Object -First 1
	if ($CliPath) {
	Copy-Item $CliPath.FullName "release\mel-band-roformer\"
	}

	# Copy DLL files
	Get-ChildItem -Path build -Filter "*.dll" -Recurse -File \| ForEach-Object {
	Copy-Item $_.FullName "release\mel-band-roformer\" -ErrorAction SilentlyContinue
	}

	# List contents
	Write-Host "Release artifact contents:"
	Get-ChildItem "release\mel-band-roformer" \| Format-Table Name, Length

	- name: Upload Release Artifact
	uses: actions/upload-artifact@v4
	with:
	name: BSRoformer-windows-cuda-${{ matrix.cuda_version }}
	path: release\mel-band-roformer\
	retention-days: 30

Navigation Menu

perf(stft): optimize FFT with table-based transforms and memory pooling #44

Workflow file

perf(stft): optimize FFT with table-based transforms and memory pooling #44

Uh oh!

Workflow file for this run