Return UI control configuration (values, limits, visibility) for model type. Args: is_turbo: Whether the model is a turbo variant.

Parse and return AST for a generation interface module. Args: module_name: Filename under ``acestep/ui/gradio/interfaces`` to parse.

Run analysis-only branches and return response payload when handled. Args: req: Generation request object. params: Prepared gener

Register the reinitialization endpoint. Inputs: app, auth dependency, response wrapper, env parser, and project-root helper. Returns: None af

Reset CUDA memory stats and free caches.

Resolve the LM backend against runtime compatibility restrictions.

Encode audio to latents using overlap-discard tiling. Args: audio: Tensor shaped ``[batch, channels, samples]`` or ``[channels, s

Determine navigation button interactive states. Returns: Tuple of ``(can_go_previous, can_go_next)`` booleans.

Build a per-sample float mask: 1.0 inside repaint, 0.0 outside, ramp at edges. Args: total_samples: Total waveform length in samples.

Create a mock model with named parameters.

Print debug message if NANOVLLM_DEBUG is enabled

Decode IPC arguments from UTF-8 JSON bytes produced by ``_encode_for_shm``. Reconstructs ``Sequence`` objects from their serialised state without

Custom CFG generation loop that: 1. Processes both conditional and unconditional sequences in parallel 2. Applies CFG formula

Get project root directory path. Returns the directory set by the ``ACESTEP_PROJECT_ROOT`` environment variable when present, otherwi

Project-local user presets directory. Anchored to the Side-Step project root rather than raw ``Path.cwd()`` so that presets are always found

Build the standard timestamp error payload.

Create a minimal torch test double for compatibility checks.

Pad sequence tensors to the same length.

Capture runtime args and return deterministic runtime state.

Return stable sample payload list for dataset endpoints.

Reproduce the handler's src_audio processing logic. This mirrors the branching in ``handler.py`` generate_music() without requiring a

Record routed args and return sentinel audio tensor.

Toggle a LyCORIS LoKr adapter via its multiplier. Args: decoder: Model decoder that may carry a ``_lycoris_net`` attribute. enabl

Reduce a ``(cond, null)``-packed prediction to a guided velocity. ``do_cfg`` is the global gate (``guidance_scale > 1``). ``apply_cfg_now`` i

EMA smoothing across *timesteps* (not across n_avg draws). Caller is responsible for invoking this once per scheduler step on the averaged ve

Write JSON atomically to reduce corruption risk during incremental saves.

Return extra format guidance inferred from the user's stated intent.

Close cache connection

Freeze all non-LoRA parameters in the model. Args: model: The model to freeze parameters for freeze_encoder: Whether to freeze th

Generate raw LM text output from a pre-built formatted prompt. Args: formatted_prompt: Prompt that is already formatted

Get parameter by canonical name from all known request sources.

Return Gradio head HTML that injects audio preference behavior. Returns: HTML snippet with a single ``<script>`` tag containing the exter

Fetch a placeholder debug mode by name.

Extract LM model size from model path. Args: model_path: Model path string (e.g., "acestep-5Hz-lm-0.6B", "acestep-5Hz-lm-0.6B-v4-fix

Get samples data in a format suitable for Gradio DataFrame.

Get parameter as integer with fallback default.

Return cached model identifiers when the entry is present and still fresh.

Load the full ``AceStepConditionGenerationModel`` for training. The model is loaded in eval mode with gradients disabled on all parameters (t

Return the parsed AST module for the provided source path.

Parse LM output to extract metadata and audio codes. Expected format: <think> bpm: 73 caption: A calm piano

Convert PMI score to normalized [0, 1] range using sigmoid function. score = sigmoid(PMI / scale) = 1 / (1 + exp(-PMI / scale)) Args:

Print a section header.

Register the ``GET /v1/audio`` route.

Return the expected system and user messages for provider request building.

Persist discovered model identifiers for a provider configuration.

Set the global GPU configuration.

Override the safe root directory. Args: root: New safe root (will be normalised and symlink-resolved).

Toggle LoRA/LoKr usage for inference.

Decode latents using tiling to reduce VRAM usage. Uses overlap-discard chunking to avoid boundary artifacts while constraining peak d

Match torch module ``to`` chaining.

Convert to dictionary for PEFT config.

Persist terminal job result payload to local cache. Args: local_cache: Cache object exposing `set(key, value, ex=...)`. store: Jo

Build request model from parsed values and explicit overrides.

Build a complete positional arg list for capture_current_params. All values default to None; supply keyword overrides for specific fields.

Build app/client pair with deterministic route dependencies.

Capture payload-builder args and return deterministic success payload.

Build state transition map based on user-provided metadata.

Build unconditional prompt for CFG based on generation phase and batch mode

Destroy default torch distributed process group when already initialized.

Collect DiT/LM inventory for status endpoints. Inputs: app state plus project/model helper callables. Returns: wrapped-internal inventory pay

Return module device string when available.

Normalize the host device value to a backend type string.

Parse boolean from environment variable.

Format seconds to VTT timestamp ``HH:MM:SS.mmm``.

Get current VRAM usage string for logging.

Load PyTorch model from path and return (success, status_message)

Build the standard lyric-score error payload.

Return a minimal LLMHandler without any loaded models.

Build a minimal trainer mock with the given decoder embedded.

Parse and normalize metadata values with safe fallbacks.

Return initialized handler or raise HTTP 500 when unavailable.

Normalise and resolve symlinks in *path*. Uses ``os.path.realpath`` so that symlinked prefixes are resolved to their canonical form before co

Run a single tier test with the given configuration. Args: sim_gb: Simulated VRAM in GB gpu_config: GPU configuration for th

Invoke ``get_effective_free_vram_gb`` with ``torch`` injected via sys.modules. The function does ``import torch`` locally, so we inject our m

Safely join user path to base root, preventing directory traversal. Uses ``os.path.normpath`` + ``startswith`` — the pattern CodeQL recognise

Sanitize a preset name to be a safe filename stem. Strips whitespace, replaces spaces with underscores, removes unsafe characters, rejects pa

Setup and configure constrained processor for generation

Check free VRAM headroom before attempting service_generate. Model weights are already resident in GPU memory at this point. We only

Replace non-repaint regions of xt with noised source latents. At each diffusion step the non-repaint regions are forced back to the appropria

Per-trajectory velocity-norm clamp. No-op when ``norm_threshold <= 0``.

Build request-model payload for ``/release_task``. Args: parser: Request parser exposing ``str``, ``bool``, ``int``, ``float``, and ``get

Return the closed timestep tensor of length ``infer_steps + 1``. Mirrors the base-variant logic: ``linspace(1, 0, n+1)`` optionally transform

Reward Model Calculator: Computes a final reward based on user priorities. Priority Logic: 1. Caption (Highest): The overall vibe/st