Extract assistant content from protocol-specific API responses.

Two-phase LM generation: CoT generation followed by audio codes generation. - infer_type='dit': Phase 1 only - generate CoT and return metas

Return raw value for ``key`` from parser payload.

Return available checkpoint directory paths under the project root.

Return deterministic labeled count.

Generate LRC timestamps by aligning decoder cross-attention to lyric tokens. Args: pred_latent (torch.Tensor): Generated latent t

Return a user-facing warning when vLLM should be skipped before initialization. Args: device: The resolved device string for LM initializ

Display an error with suggestions. Args: exc: The exception that occurred. context: Short label (e.g. "Model loading", "Training"

Auto-fill audio duration from source audio in Extract/Lego mode. Reads duration directly via soundfile to avoid the training-module safe_path

Initialize LLM model according to GPU config and environment overrides.

Ensure app state has a stable ``training_state`` mapping.

Return whether lyrics indicate instrumental output.

Load the full persisted provider-settings payload.

Load persisted external LM settings for a provider or the active selection.

Materialize non-parameter MLX buffers before cross-thread use.

Move decoder to CPU to free VRAM for non-decoder workloads.

Double encoder/context/attention tensors for CFG, or pass through. Returns a 4-tuple ``(enc_hs, enc_am, ctx, attn)`` ready for one ``decoder(

Parse language and instrumental hints from free-form description text. Args: description: User free-form sample description. Returns

Prepare per-item seeds and UI seed string.

Load and normalize reference audio, then sample 3x10s segments. Args: audio_file: Path to reference audio file. Returns:

Return encoded payload bytes.

Parse CLI args, set environment overrides, and start uvicorn. Args: env_bool: Boolean env parser used for defaults. argv: Optiona

Run music generation and aggregate outputs across sequential MPS cover slices. Args: req: Generation request object. job_id: Job

Execute one queued job end-to-end with success/failure cache updates. Args: app_state: FastAPI app state object. store: Job store

从 base64 data URL 保存音频文件

Save LoKr weights to safetensors.

Persist uploaded audio file to a temporary location. Args: upload: Uploaded file wrapper from Starlette/FastAPI. prefix: Filename

Resolve the handler/model name for a generation job request. Args: app_state: Application state object containing primary/secondary/third

Record service invocation and return minimal output payload.

Set the instance-level default language (shared fallback when no ContextVar is active).

Setup datasets.

(Re)start TensorBoard with the given logdir and return URL if successful.

Coerce optional float values.

Understand audio codes and generate metadata + lyrics. This is the reverse of the normal generation flow: - Input: Audio cod

Understand music from audio codes using the 5Hz Language Model. This function analyzes audio semantic codes and generates metadata about the

Unload LLM to release VRAM and mark state flags.

Move models to CPU, delete references, and free GPU memory. Accepts any number of model objects (or ``None`` values, which are silently skipp

Best-effort unwrap for common wrapper attributes used by training runtimes.

Record update increments from mixin helpers.

Update progress and optional stage for an existing job.

Validate Authorization header for endpoints that do not parse request body. Args: authorization: Optional ``Authorization`` header value.

Verify API key from request body (ai_token) or Authorization header. Returns the token if valid, None if no auth required.

Singleton pattern

Add arguments shared by vanilla / fixed subcommands.

Normalize and add a model ID from a response item.

Apply runtime hardware overrides for LM backend selection.

Apply top-k filtering to logits

Apply top-p (nucleus) filtering to logits

Inverse of :func:`_btc_to_bct`.

Create app state and register sample/format routes for tests.

Replicate the audio_params building logic from inference.py.

Create app/client pair with scan/load routes only.

Create app/client pair with lightweight dataset + model state.

Create app and register sample/format routes.

Build the composite Rich renderable for one Live refresh.

Build logits processor list with repetition penalty if needed

Build ordered candidate model-list URLs from a provider endpoint.

Shipped built-in presets directory.

Remove temporary directory silently.

Release GPU memory between session-loop iterations.

Apply backend constraints and return normalized compile/quantization settings.

Move VAE to CPU, decode there, then restore original device.

Run DWT on both latents. Returns ``(xl, xh, yl, yh, iwt, out_T)`` or ``None`` if the optional ``pytorch_wavelets`` dependency is missing.

Release cached GPU memory (CUDA / MPS / XPU).

Force trainable floating-point parameters to fp32.

Run shared format-sample workflow. Returns: Tuple of ``(result_or_none, audio_duration_value_or_none, status_message)``.

Extract lyrics section from LLM output. The lyrics appear after the </think> tag and typically start with "# Lyric" or direc

Return the fixed 46 core UI outputs from a generation result tuple. The generate-button wiring expects 46 generation outputs from the wrapper,

Return a fake keyring module with a KeyringError base class.

Build a standardized failure response with update placeholders.

Ensure instruction ends with a colon.

Format lyrics text with language header.

Return the AffineQuantizedTensor class from torchao, or None if unavailable.

Return endpoint callable matching a route path/method pair.

Return endpoint callable matching route path and method.

Get the trie node for a given prefix. Returns None if the prefix is not valid (no genres start with this prefix).

Convert internal model name to OpenRouter model ID.

Extract model name from config path.

Check if negative prompt is meaningful (not default/empty)

Import service_init lazily to avoid heavy transitive imports.

Stub MLX DiT init hook and always report unavailable in tests.

Return whether a device identifier refers to any CUDA device.

Load the full cache payload from disk when available.

Load cached repaint source state from a generated audio sidecar. The cache is an optimization for ACE-generated Gradio outputs. Missing or ma