QVAC-23802 feat[audiogen-ggml]: add an opt-in cuda feature

[Zbig9000]

Status

Ready. Final link of QVAC-23802 (AudioGen CUDA for ACE-Step and MiniMax). One commit, 19 added lines, no behaviour change to any existing build.

Why

The engine work merged in tetherto/qvac-ext-lib-whisper.cpp#162 and the ports in tetherto/qvac-registry-vcpkg#326 (which added speech-cpp's cuda feature — it previously exposed only metal / vulkan / opencl). This addon was the remaining gap: its manifest hardcoded speech-cpp[audiogen,vulkan] for desktop and exposed only ENABLE_VULKAN, so a CUDA build was not selectable from here and the merged CUDA support stayed unreachable for consumers.

What changes

Mirrors the existing vulkan plumbing exactly:

• CMakeLists.txt: option(ENABLE_CUDA ...) appending the cuda manifest feature.
• vcpkg.json: a cuda feature depending on speech-cpp[cuda], which in turn pulls ggml-speech[cuda].

Off by default, and excluded on Apple and Android to match speech-cpp's own supports: "!osx & !ios & !android", so every current build resolves exactly as before. nvcc is needed only on the build host; the runtime needs just the driver. No CUDA-specific staging was required — the backend install plumbing already iterates GGML_AVAILABLE_BACKENDS, so a CUDA-enabled ggml is picked up by the existing code.

Verification

vcpkg install --dry-run --x-feature=cuda --triplet=x64-linux against this manifest resolves the whole chain:

ggml-speech[core,cuda,vulkan]:x64-linux@2026-08-21
speech-cpp[audiogen,core,cuda,vulkan]:x64-linux@2026-08-24#2

CUDA is additive alongside Vulkan, and the engine's validated-GPU preference selects CUDA when both backends are compiled in. A control run with a deliberately bogus --x-feature=cudaXX warns not a feature supported, which the cuda run does not — confirming the feature is recognised rather than silently ignored.

Engine-level CUDA behaviour behind this feature was validated on an RTX 5090 (CUDA 12.9, sm_120) while landing #162 / #326: ACE-Step test-acestep-integration passes on CUDA and CPU, MiniMax test-minimax-quality passes with MM3_DEVICE=gpu, and the full speech superbuild compiles with GGML_CUDA=ON (661/661 targets, ctest -L unit 62/62).

Out of scope

Prebuilds stay Vulkan-only. The reusable prebuild workflow provisions a Vulkan SDK (include-vulkan-sdk: true) and has no CUDA equivalent, so shipping CUDA artifacts needs nvcc on the builders — shared-CI-infra work rather than a change to this package. This PR only makes a CUDA build selectable.

Note this is the first ENABLE_CUDA lane in the monorepo; tts-ggml merged three CUDA engine tickets (Chatterbox, Supertonic, Parler-TTS) and would want the identical two-hunk change if the platform owners want to generalise it.

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1217591858523925

[github-actions]

Review Status

Current Status: ❌ PENDING
Approvals so far: none

Pending reviews: Needs 1 Management or Team Lead, and 1 more from Management, Team Lead, or Member.

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./.github/actions/release-merge-guard
• ./docs/website
• ./packages/ggml-coload-smoke
• ./packages/fabric/test/integration
• ./packages/inference-addon-cpp/mobile
• ./packages/sdk/e2e
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/vla-ggml/sim/server
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/asr-ggml/benchmarks/server