Capture current UI parameters for next-batch generation. Audio codes are cleared so AutoGen batches always generate fresh content.

Check if the main model components exist in the checkpoints directory. Returns: True if all main model components contain weights, F

Check if a specific model exists in the checkpoints directory. Args: model_name: Name of the model to check checkpoints_

Create a music sample from a natural language query using the 5Hz Language Model. This is the "Simple Mode" / "Inspiration Mode" feature that

Download the main ACE-Step model from HuggingFace or ModelScope. The main model includes: - acestep-v15-turbo (default DiT model) -

Dynamic Time Warping algorithm optimized with Numba. Args: x: Cost matrix of shape [N, M] Returns: Tuple of (text_indic

Ensure the main model is available, downloading if necessary. This function is designed to be called during initialization. It will only

Evaluate legacy GPU compatibility using an injected torch-like module. Args: torch_module: Imported ``torch`` module or a test double exp

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

Return DCW defaults for the current Think state. Args: think_enabled: Whether LM Think mode is enabled. Returns: Defaults fo

Return the provider profile for a provider identifier. Raises: ValueError: If ``provider`` is not a supported external provider.

Get GPU memory in GB. Returns 0 if no GPU is available. Debug Mode: Set environment variable MAX_CUDA_VRAM to override the detected

Get preview data for a specific sample. Returns: Tuple of (audio_path, filename, caption, genre, prompt_override, lyrics,

Check if sample has user-provided raw lyrics from .txt file.

Infer packed reference-audio latents and order mask.

Detect GPU configuration and optionally initialize models at startup. By default models are NOT loaded at startup (lazy-loaded on first request).

Check if running on macOS with MPS (Apple Silicon) available. This is the canonical check used across the codebase to apply Mac-specific conf

Return whether this host simulates a turbo model.

Parse and return AST for an interfaces module.

Load the best JSON object candidate from provider text content.

Load a preprocessed tensor dataset for training. Returns: Info text about the dataset.

Log a single scalar value.

Apply median filter to tensor. Args: x: Input tensor filter_width: Width of median filter Returns: Filtered ten

Parse ``/release_task`` request body into request model and temp-file list. Args: request: FastAPI request carrying body/form data.

Persist per-sample intermediate tensors beside an audio JSON sidecar. Args: extra_outputs: Generation result extra outputs containing bat

Preprocess dataset to tensor files for fast training. Converts audio files to VAE latents and text to embeddings. Returns: Status me

Load and normalize source audio for remix/extract flows. Args: audio_file: Path to source audio file. Returns:

Read adapter-specific value from mapping-like or scalar containers.

Resolve the Remix ``no_fsq`` checkbox into the backend task type. Args: task_type: Current hidden Gradio task type. no_fsq: Wheth

Map JS restore results into Gradio output values. The JS function reads localStorage and produces an array: - First ``len(PREF_KEYS)`` elem

Save LoRA adapter weights. Args: model: Model with LoRA adapters output_dir: Directory to save weights save_full_model: W

Set the API key for authentication.

Display an info message.

Temporarily switch LLM model for a critical section and restore afterward.

Train LoRA adapters from preprocessed tensor files. This is the recommended training method for best performance. Args:

Single training step using preprocessed tensors. Note: This is a distilled turbo model, NO CFG is used. Args: batch: Dic

Update UI settings based on model type (fallback when handler not initialized yet). Args: config_path: Model config path string.

Initialize deterministic runtime state for decode tests.

Apply temperature scaling based on current generation phase. Temperature scaling: logits = logits / temperature - Lo

Create test app and register LoRA routes with stubs.

Create app state and register query-result route for tests.

Build app and register release-task route with deterministic fakes.

Build a FastAPI app, register routes, and return an HTTP client.

Build prefix tree for numeric field based on actual tokenization with context. IMPORTANT: Uses token ID sequences as keys, N

Calculate average log probability of target text given prompt.

Clean up handlers and free GPU memory.

Return whether a model directory contains at least one weights artifact. Args: model_path: Candidate model directory path. Returns:

Return the CUDA device index implied by a device identifier.

Print debug message if NANOVLLM_DEBUG is enabled

Capture decode args and return deterministic decode outputs.

Read the current API key under lock.

Return the endpoint callable matching ``path`` and HTTP ``method``.

Return endpoint callable matching a route path/method pair.

Get nested dictionary value using dot notation Args: data: Dictionary to search key: Dot-separated k

Resolve route endpoint callable by path and HTTP method.

Return current process RSS in megabytes. Uses ``/proc/self/statm`` on Linux for the true current resident set size. Uses ``ctypes`` o

Return a materialized sliding-window mask for the requested sequence length.

Platform-aware global user presets directory (fallback). Used as a secondary scan location so that presets saved before the local-directory c

Create a fake GPU config object with all fields used by startup init.

Single-level Haar DWT along the T axis of a ``[B, T, C]`` ``mx.array``. Returns ``(low, high)``, each of shape ``[B, T//2, C]``. If ``T`` is

Return ``True`` when at least one non-empty audio-code string is present.

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

Return True if the raw input string is a back-navigation request.

Check whether backend operation result is considered successful.

Load cached repaint source latents from a generated audio sidecar.

Return a minimal completed batch dict for testing.

Build a minimal object that satisfies the mixin's interface.

Return a mock that looks like a PEFT PeftModel-wrapped decoder.

Build a ``torch`` mock whose ``.cuda`` reports deterministic memory stats. Args: device_free_bytes: Bytes free at the CUDA-driver level.

Return deterministic max-memory value for debug logging.

Encode one MLX audio sample with optional overlap-discard tiling. Args: audio_nlc: MLX array in ``[1, samples, channels]`` layout

Return a unique DOM id suffix for each help modal instance.

Normalize audio tensor to stereo at 48kHz. Args: audio: Tensor in [channels, samples] or [samples] format. sr: Source

Move the auxiliary HF scoring model off accelerator memory after PMI.

Read a launcher script from the project root.

Resolve a concrete runtime device, applying backend fallback rules.

Execute OpenSSL with the passphrase stored in a temporary file.

Build deterministic tensors and lyric token ids for tests.

Synchronize accelerator operations (CUDA, XPU, or MPS).

Blend generated latents with source at repaint boundaries. Args: x_gen: Diffusion output [B, T, C]. clean_src_latents: Original s

Build the format-mode user intent string for external provider requests.

Extract a simple call-target name from an AST call function node. Args: node: AST node representing a call target expression. Return

Mark fake upload as closed.

Compute normalized defaults shared by generation interface sections. Args: init_params: Optional startup parameters passed to the UI laye

Convert uploaded source audio into serialized audio code tokens.

Create and register a queued job using a caller-provided ID.

Download a specific sub-model from HuggingFace or ModelScope. Args: model_name: Name of the model to download (must be in SUBMODEL_R

Ensure models are initialized, loading them lazily on first request if needed. If models were already loaded at startup (``ACESTEP_NO_INIT=false`

Return the user-local JSON cache path for discovered provider models.