Backend reference

Backends

Support matrix

Note: The llamacpp rocm row lists linux, windows for the family as a whole, but MI350X (gfx950) is currently gated to Linux + stable channel only — the Windows TheRock distribution and the ROCm nightly build for gfx950 are not yet published, so gfx950 installs are rejected on Windows and on the nightly channel. The OS column reflects the row's overall reach; the per-architecture restriction is enforced by the backend's install gate.

Recipe options

acestep — ACE-Step

flm — FastFlowLM NPU

llamacpp — Llama.cpp GPU

moonshine — Moonshine

onnxruntime — ONNX Runtime

openmoss — OpenMOSS TTS

sd-cpp — StableDiffusion.cpp

thenoise — TheNoise ROCm (experimental)

thinksound — ThinkSound

trellis — TRELLIS.2

vllm — vLLM ROCm (experimental)

whispercpp — Whisper.cpp

Implementation notes

ACE-Step (acestep)

ace-server exposes an asynchronous job API: POST /lm or POST /synth returns a job id immediately, GET /job?id=N polls the status, and GET /job?id=N&result=1 fetches the finished result. AceStepServer::run_job wraps this submit/poll/fetch cycle with a ceiling of roughly 20 minutes per stage at a 1-second poll cadence. Synth results arrive as multipart/mixed (an audio part plus a latent part); Lemonade extracts the first audio part.

Vocals are a two-stage pipeline. POST /lm with lm_mode: "generate" runs the ACE-Step language model, which turns the caption and lyrics into audio codes plus LM-filled metadata, returned as a JSON array of enriched requests. That array is accepted by POST /synth verbatim, so Lemonade feeds it through unchanged. POST /synth on its own is the DiT-only instrumental path — it has no language model and cannot sing, so a lyrics value (other than the sentinel below) is what routes a request through /lm first.

[Instrumental] — any case, surrounding whitespace ignored — is ACE-Step's sentinel for the no-vocals path, matching the Python reference implementation. Instrumental requests send the sentinel explicitly rather than an empty string because the synth stage also feeds the lyrics text into its conditioning.

Errors from audio_generations are written into the response sink as a JSON error payload; the endpoint handler turns that into an HTTP error instead of shipping it as audio.

The model download fetches the DiT checkpoint variant plus three companions when present in the repo: the language model (acestep-5Hz-lm-4B-Q8_0.gguf, required for vocals and auto-lyrics), the Qwen3 text encoder, and the VAE. The checkpoint path handed to --models is the directory of GGUFs; ace-server scans it by architecture, and --keep-loaded keeps models resident across requests.

Backend auto-selection

When a recipe's backend is not pinned in config.json (the backend key is absent or "auto"), the default backend reported in system-info — and used by RecipeOptions when resolving *_backend options — is chosen as follows: the first supported backend in RECIPE_DEFS preference order wins, unless a later supported backend is already locally installed (state installed, update_available, or update_required) while the earlier candidates are merely installable. In that case the first installed one wins. This makes explicitly installing a variant (e.g. the Vulkan build of a GPU backend) an effective override: auto-selection uses what is on disk instead of downloading the preference-order winner. An explicit backend value in config.json always takes precedence over both rules. The llamacpp system variant is never auto-selected unless prefer_system is set.