Return the Python keyring module when available.

Load the external save-preferences JavaScript asset.

Return a mock constrained processor pinned to *state*.

Return a mock that looks like a Fabric _FabricModule wrapping *inner*.

Create a KV cache for the MLX model.

Return a minimal host configured as if running on ROCm.

Return deterministic decoded waveform for MLX decode branch.

Move an AffineQuantizedTensor to target device using ``_apply_fn_to_data`` when available.

Normalize batch input: convert single string to list and return (list, is_batch)

Convert numpy audio array to channels-first torch tensor.

Right-pad/copy ``values`` to ``batch_size``, falling back to ``default``.

Parse a Python source file in the CLI directory and return its AST. Args: filename: Name of the file relative to the CLI package director

Prepare model-ready conditioning batch tensors and metadata. Args: captions: Per-item captions. lyrics: Per-item lyri

Capture decode-state args and return deterministic latents/costs.

Aggressively reclaim system memory after device transfers. Combines Python GC, accelerator cache flush, and OS-level heap trimming to

Advance the generator to its first yield and return the tuple.

Call handler.initialize() with all heavy operations mocked. Args: flash_attn_available: Whether flash_attn should appear detectab

Run ``main`` with heavy dependencies stubbed and capture startup state.

Run the two-pass preprocessing pipeline from the standalone entrypoint. Args: args: Parsed CLI arguments with ``audio_dir``, ``tensor_out

Sample tokens from logits with temperature. Upcasts to float32 for numerical stability (float16 logits can overflow during softmax, e

Save MP3 with explicit ffmpeg settings and 128k/48k defaults.

Smart download with automatic fallback between HuggingFace and ModelScope. Automatically detects network environment and chooses the best do

Copy .py files from acestep/models/{variant}/ into the checkpoint directory, overwriting the HuggingFace-downloaded versions. Args:

Temporarily unload interactive vLLM so PMI can use HF scoring model. nano-vLLM does not support a cheap ``.to("cpu")`` offload. For PMI scor

Tiled encode that offloads latent chunks to CPU immediately.

Compute the DCW correction on MLX arrays via pytorch_wavelets. One numpy copy per latent; runs on CPU with float32. Returned as ``mx.array``

Unwrap stale Lightning Fabric wrappers from decoder left by previous runs. Returns: True if decoder was unwrapped, else False.

Return whether the current platform should prefer the system keyring.

Validate quantization prerequisites before model loading.

Forward to lifecycle implementation to mimic mixin wiring.

Append a single JSONL record for audit/progress tracing.

Apply scale to one adapter and return `(modified_count, report)`.

Build the chat-formatted prompt for audio understanding from codes. This is the reverse of generation: given audio codes, generate m

Extract a simple call-target name from an AST call function node.

Return whether the generated caption looks like a non-enhanced echo.

Check if LyCORIS is importable.

Check if PEFT library is available.

Enable gradient checkpointing, disable use_cache, and enable input_require_grads across all wrapper layers of *decoder*. Mirrors ACE-Step's `

Return configured conversion output.

Create silent stereo target audio with safe duration handling.

Ensure a DiT model is available, downloading if necessary. Args: model_name: Name of the DiT model checkpoints_dir: Custom c

Ensure a mock gradio module exists for package imports in tests.

Ensure an LM model is available, downloading if necessary. Args: model_name: Name of the LM model (defaults to DEFAULT_LM_MODEL)

Match torch module ``eval`` API.

Return the user-local JSON settings path for external LM preferences.

Extract balanced top-level JSON object candidates.

Extract a likely JSON object from plain or fenced content.

Find the best matching disk model for a recommended tier model. If the exact recommended model exists on disk, return it. Otherwise, fin

No-op flush to satisfy file-like API expectations.

Return generated audio codes from a JSON sidecar when available.

Get caption with custom tag applied.

构建请求头

Get the number of labeled samples.

获取延迟统计数据

Calculate lyric alignment scores for pure-noise and regressed-latent inputs. Args: pred_latent (torch.Tensor): Generated latent t

处理 API 响应并保存音频

Infer text-token embeddings via text encoder.

Initialize process environment, model handlers, and app.state runtime fields.

Remove matching cached model entries for a provider configuration.

Treat near-zero tensors as silence.

Capture label invocation arguments and return success status.

Return launcher probe exit code for legacy torch compatibility check. Args: torch_module: Optional injected ``torch`` module for tests.

Load artifact tensors for a one-based sample index in a batch queue entry.

Load a random example from the task-specific examples directory. Args: task_type: The task type (e.g., "text2music"). llm_handler

Convert LRC text to a VTT subtitle file and return its path. Args: lrc_text: LRC format lyrics string. total_duration: Total audi

Mark a job as failed and persist an error message.

Apply ``torch.compile`` only when its backend (Triton) is available. Drop-in replacement for the ``@torch.compile`` decorator. When Triton i

Cached check for MLX availability.

Navigate to the next batch (result view only). Yields: Two tuples of 49 Gradio component updates each.

Validate and normalize a supported external AI protocol identifier.

Parse comma-separated timesteps into float list.

Calculate Pointwise Mutual Information (PMI) score. PMI = log P(condition|codes) - log P(condition) = log [P(codes|condition) / P(co

Resolve LLM flags and prepare caption/lyrics/metadata inputs for generation. Args: app_state: FastAPI app state with LLM initialization f

Process one queued job item and notify waiters. Args: job_id: Job identifier from queue. req: Request payload associated with the

Register LoRA lifecycle endpoints on the provided FastAPI app.

Register health/stats/inventory/init routes. Inputs: app, auth/response wrappers, job store settings, and model helper callables. Returns: No

Register the ``/query_result`` endpoint. Args: app: FastAPI app instance to register the route on. verify_token_from_request: Aut

Register random-sample and format-input routes on the FastAPI app.

Reset to pre-step state and advance once with the averaged diff. ``avg_diff_*`` is ``None`` when CFG was inactive this step (no averaged diff