The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.
Rust ffmpegfree-video-editormodel-context-protocolnon-linear-editoroffline-firstopen-source-video-editorself-hostedvideo-editingvideo-editing-automationvideo-editing-softwarevideo-editing-toolsvideo-editorvoice-cloningyoutube-automationyoutube-shorts
Fetch sherpa-onnx outside target/, key the cache on the build environment, and build ggml portable
Jareerc6d79535f5e5
diff --git a/.github/workflows/build-app.yml b/.github/workflows/build-app.ymlindex b1a6428c..79c116a4 100644--- a/.github/workflows/build-app.yml+++ b/.github/workflows/build-app.yml@@ -55,6 +55,13 @@ env:# Oldest macOS the bundles run on. Apple silicon Macs all run newer than# this; it is the Intel floor.MACOSX_DEPLOYMENT_TARGET: "12.0"+ # ggml (whisper.cpp's kernels) tunes itself to the machine it is built+ # on unless told otherwise, and these bundles run on other machines: a+ # runner with AVX-512 or SVE would leave a binary that faults on a CPU+ # without them. Off, ggml compiles to the portable baseline instead -+ # AVX2 with FMA and F16C on x86_64, NEON on ARM. whisper-rs-sys forwards+ # every GGML_* variable to CMake as a define.+ GGML_NATIVE: "OFF"jobs:app:@@ -63,14 +70,18 @@ jobs:fail-fast: falsematrix:include:+ # `ffmpeg` and `sherpa` name the target the way each project's+ # release archives do.- platform: macosarch: arm64os: macos-15target: aarch64-apple-darwin+ sherpa: osx-arm64- platform: macosarch: x86_64os: macos-15-inteltarget: x86_64-apple-darwin+ sherpa: osx-x64# 22.04, not latest: the binary links the build machine's glibc,# so the oldest supported runner is what keeps it runnable on# older distributions.@@ -79,21 +90,25 @@ jobs:os: ubuntu-22.04target: x86_64-unknown-linux-gnuffmpeg: linux64+ sherpa: linux-x64- platform: linuxarch: aarch64os: ubuntu-22.04-armtarget: aarch64-unknown-linux-gnuffmpeg: linuxarm64+ sherpa: linux-aarch64- platform: windowsarch: x86_64os: windows-2022target: x86_64-pc-windows-msvcffmpeg: win64+ sherpa: win-x64- platform: windowsarch: aarch64os: windows-11-armtarget: aarch64-pc-windows-msvcffmpeg: winarm64+ sherpa: win-arm64runs-on: ${{ matrix.os }}# The slowest target (Intel macOS, fat LTO) finishes in under half an# hour. A job still running past this is stuck, not slow.@@ -101,10 +116,6 @@ jobs:steps:- uses: actions/checkout@v7- uses: dtolnay/rust-toolchain@stable- - uses: Swatinem/rust-cache@v2- with:- workspaces: engine- key: ${{ matrix.target }}- name: Read the versionid: version@@ -179,7 +190,12 @@ jobs:# rest of the build; the cc crate hands it the Visual Studio# environment, and the cmake crate passes it to CMake - but only# under a Ninja generator, since the Visual Studio generator chooses- # the compiler itself. Ninja ships on the runner image.+ # the compiler itself. Ninja ships on the runner image. Under Ninja+ # the cmake crate also writes CMAKE_CXX_FLAGS itself, in place of+ # CMake's defaults, and those defaults are where /EHsc came from;+ # clang-cl, unlike cl, rejects `try` outright without it (ggml's+ # gguf reader uses exceptions). CXXFLAGS travels through the cc+ # crate into those flags.- name: FFmpeg development libraries (Windows)if: matrix.platform == 'windows'shell: pwsh@@ -191,6 +207,7 @@ jobs:Start-Process -Wait -FilePath .\llvm.exe -ArgumentList "/S""CC=clang-cl" | Out-File -FilePath $env:GITHUB_ENV -Append"CXX=clang-cl" | Out-File -FilePath $env:GITHUB_ENV -Append+ "CXXFLAGS=/EHsc" | Out-File -FilePath $env:GITHUB_ENV -Append"CMAKE_GENERATOR=Ninja" | Out-File -FilePath $env:GITHUB_ENV -Append} else {choco install llvm -y --no-progress@@ -204,14 +221,39 @@ jobs:"FFMPEG_DIR=$dir" | Out-File -FilePath $env:GITHUB_ENV -Append"$dir\bin" | Out-File -FilePath $env:GITHUB_PATH -Append- # sherpa-onnx links shared on Windows (see concat-speech's manifest for- # why static cannot work here), from a directory holding only the- # import libraries. The separation matters: when the sys crate's build- # script sees DLLs beside the .lib files it tries to copy them into- # the "profile directory", located by searching OUT_DIR's ancestors- # for one named after $PROFILE - which panics under the custom `app`- # profile, whose directory is not called "release". The DLLs are kept- # apart and the packaging step ships them beside the binary.+ # ── sherpa-onnx ────────────────────────────────────────────────────+ # The prebuilt library the sys crate would otherwise fetch for+ # itself, fetched here instead and handed over through+ # SHERPA_ONNX_LIB_DIR. Left to the crate, the download lands inside+ # target/, where rust-cache's cleanup does not keep it; a restored+ # cache then replays the build script's answer without the files+ # it points at, and the crate fails to link the moment anything+ # makes it recompile. Outside target/, the files are there on every+ # run. The version is the one the lockfile resolved, so a bump+ # cannot desynchronise.+ - name: sherpa-onnx static libraries (macOS and Linux)+ if: matrix.platform != 'windows'+ shell: bash+ run: |+ set -euo pipefail+ version=$(grep -A1 'name = "sherpa-onnx-sys"' engine/Cargo.lock | sed -n 's/^version = "\(.*\)"/\1/p')+ stem="sherpa-onnx-v$version-${{ matrix.sherpa }}-static-lib"+ echo "Downloading $stem"+ curl -fsSL --retry 5 --retry-all-errors -o sherpa.tar.bz2 "https://github.com/k2-fsa/sherpa-onnx/releases/download/v$version/$stem.tar.bz2"+ mkdir -p sherpa+ tar -xjf sherpa.tar.bz2 -C sherpa+ test -d "sherpa/$stem/lib"+ echo "SHERPA_ONNX_LIB_DIR=$PWD/sherpa/$stem/lib" >> "$GITHUB_ENV"++ # On Windows it links shared (see concat-speech's manifest for why+ # static cannot work there), from a directory holding only the+ # import libraries. The separation matters: when the sys crate's+ # build script sees DLLs beside the .lib files it tries to copy them+ # into the "profile directory", located by searching OUT_DIR's+ # ancestors for one named after $PROFILE - which panics under the+ # custom `app` profile, whose directory is not called "release". The+ # DLLs are kept apart and the packaging step ships them beside the+ # binary.## The archive is .tar.bz2, and Windows' own tar has no bzip2 of its# own: it pipes through an external decoder, and on the x64 image@@ -223,10 +265,8 @@ jobs:shell: pwshrun: |$ErrorActionPreference = "Stop"- # The version the lockfile resolved, so a bump cannot desynchronise.$version = (Select-String -Path engine\Cargo.lock -Pattern 'name = "sherpa-onnx-sys"' -Context 0,1).Context.PostContext[0] -replace '.*"(.*)".*', '$1'- $arch = if ("${{ matrix.arch }}" -eq "aarch64") { "win-arm64" } else { "win-x64" }- $stem = "sherpa-onnx-v$version-$arch-shared-MT-Release-lib"+ $stem = "sherpa-onnx-v$version-${{ matrix.sherpa }}-shared-MT-Release-lib"Write-Host "Downloading $stem"curl.exe -fsSL --retry 5 --retry-all-errors -o sherpa.tar.bz2 "https://github.com/k2-fsa/sherpa-onnx/releases/download/v$version/$stem.tar.bz2"Write-Host "Unpacking"@@ -239,6 +279,17 @@ jobs:"SHERPA_DLL_DIR=$PWD\sherpa-dlls" | Out-File -FilePath $env:GITHUB_ENV -Append# ── Build ──────────────────────────────────────────────────────────+ # The cache comes after the environment is set, not before: its key+ # is made from the compiler and linker variables in force (RUSTFLAGS,+ # CC, CXX, CMAKE_*, and GGML_* added to the default set), so a change+ # to any of them is a new key rather than a stale target/ served+ # under the old one.+ - uses: Swatinem/rust-cache@v2+ with:+ workspaces: engine+ key: ${{ matrix.target }}+ env-vars: CARGO CC CFLAGS CXX CMAKE RUST GGML+# The shipping profile: fat LTO, panic=abort, stripped. See the# workspace Cargo.toml for what each knob costs.#