Print debug message if NANOVLLM_DEBUG is enabled

Create a minimal fake ``mlx.core`` module backed by NumPy arrays.

Build a training/validation loss plot (matplotlib Figure) for ``gr.Plot``.

Extract a simple function name from a call node target.

General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode. Supports use cases: 1. Causal Full: is_causal=True, is_s

General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode. Supports use cases: 1. Causal Full: is_causal=True, is_s

General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode. Supports use cases: 1. Causal Full: is_causal=True, is_s

General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode. Supports use cases: 1. Causal Full: is_causal=True, is_s

General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode. Supports use cases: 1. Causal Full: is_causal=True, is_s

General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode. Supports use cases: 1. Causal Full: is_causal=True, is_s

Return ``("ok", True)`` to emulate successful LLM initialization.

Initialize DiT and optional LM models. Args: app_state: FastAPI application state carrying handler objects and flags. model_name:

Parse LRC lyrics text to Gradio subtitles format with smart post-processing. Merges lines that start very close to each other so they don't disap

Parse ``--quantization`` values from CLI input. Args: value: Raw CLI value. Returns: Canonical quantization method or ``None

Store batch results in queue with all generation parameters. Args: batch_queue: The mutable batch queue dict. batch_index: Index

Build the client-side JS that reads localStorage and returns values. The returned function is passed as the ``js`` parameter to ``demo.load()

Release freed accelerator memory back to the driver. Synchronises the device *before* releasing cached blocks so that every in-flight

Build a _Host with a None decoder for config-only tests.

Return deterministic allocated-memory value for debug logging.

Create a slider with a non-zero range to avoid Gradio math errors.

Wrap response data in standard format compatible with CustomAceStep.

ADG (Angle-based Dynamic Guidance) forward pass for Flow Matching. In flow matching (including SD3), sigma represents the current timestep t

Build an audio slot update that always rewinds playback to 0. Args: gr_module: Module/object exposing ``update(**kwargs)``. audio

Return deterministic form payload for this test request.

Generate deterministic UUID from generation parameters. Same parameters will always generate the same UUID. Args: params_dic

Get global i18n instance Args: language: Optional language to set Returns: I18n instance

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

Return Gradio head HTML that injects save-side preference persistence.

Initialize 5Hz LM model Args: checkpoint_dir: Checkpoint directory path lm_model_path: LM model path (relati

Check whether a model path refers to a pure base model. Args: config_path_lower: Lowercased model config path string. Returns:

Return precomputed latent sample.

Return a tiny state dict suitable for backup/restore paths.

Load model_downloader directly without importing heavy dependencies.

Build SSE chunk JSON string.

Recursively move parameters and buffers to the specified device. Tries the fast ``model.to()`` path first. Only falls back to manual

Convert repaint mode and strength into concrete numeric parameters. Higher *strength* means more aggressive repainting (less source preservation)

Transition to the next FSM state.

Splice original user waveform into non-repaint regions after VAE decode. Args: pred_wavs: VAE-decoded waveform ``[B, C, samples]``.

Emit a timestamped debug log line if the mode is enabled.

Run paired flow-edit on top of text2music dispatch. Builds source text/lyric embeddings from ``flow_edit_ctx['source_caption']`` / ``['source

构建请求头

Check whether flash attention can be used on the target device.

Pretend to restore weights and report no key mismatches.

Convert a path string to use the native OS separator for display. On Windows, replaces forward slashes with backslashes so that path defaults

Set the per-request language for the current execution context. Call at a request boundary (Gradio event handler, FastAPI middleware) to scop

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

Compute max_new_tokens based on target duration and generation phase. In the two-phase architecture: - CoT phase: generates

Encode IPC arguments to UTF-8 JSON bytes for shared-memory transport. Replaces ``pickle`` to avoid arbitrary code execution on deserialization.

Return a silence-latent slice of exactly ``length`` frames. When the pre-computed ``silence_latent`` tensor is shorter than ``length`

Run the MLX diffusion loop and return generated latents. Args: encoder_hidden_states: Prompt conditioning tensor. enc

Capture audio-prepare args and return deterministic prepared state.

Build a soft float mask with linear crossfade at repaint boundaries. The crossfade zones extend into the preserved (non-repaint) region on each

End timing only in VERBOSE mode for a named subsystem.

Emit a timestamped debug log only in VERBOSE mode for a named subsystem.

Start timing only in VERBOSE mode for a named subsystem.

Detect the best available accelerator. Priority: CUDA > MPS > XPU > CPU. Args: requested_device: 'auto', 'cuda', 'cuda:N', 'mps', 'x

Run the shared flow-edit sampling loop. See module docstring for the algorithm. ``model`` is the DiT variant (must expose ``.decoder(...)``).

Format user-provided caption and lyrics using the 5Hz Language Model. This function takes user input (caption and lyrics) and generates struc

Generate audio from text/reference inputs and return response payload. Args: captions: Text prompt describing requested music.

Scan and return all model directory names starting with ``acestep-v15-``.

Check whether a model path refers to an SFT (supervised fine-tuned) model. Args: config_path_lower: Lowercased model config path string.

Map textual job status to integer API status code.

Restore the per-request language to its previous value.

Restore previously offloaded components back to their original state.

Save adapter weights directly into *output_dir* (no nesting). Writes ``adapter_config.json`` and ``adapter_model.safetensors`` (or LoKR equiv

Set the process-level API key used by auth verification helpers. Args: key: API key string, or ``None`` to disable auth enforcement.

Reinitialize with quantization disabled using the last successful init parameters. Returns: Tuple[str, bool]: - A

Translate *key* using per-request ContextVar, then instance default, then English.

Run the full training loop. Yields ``(global_step, loss, status_message)`` tuples.

Build a compact generation timing summary. Args: lm_metadata: LM-generated metadata dictionary (unused, kept for API compat). tim

Get the next field state. Always returns the next field's NAME state, even if the field is user-provided (we still need to generate t

Decode a single MLX latent sample with optional tiling. Args: z_nlc: MLX array in ``[1, frames, channels]`` layout. d

Move a module to a target device/dtype when possible.

Read an audio file, with torchaudio fallback for formats unsupported by soundfile. soundfile (libsndfile) supports WAV/FLAC/OGG/MP3 etc. but NOT

Map precision string to torch dtype.

Reference implementation mirroring the inline mix inside generate_audio.

Unwrap PEFT/Fabric wrappers from a model/decoder to retrieve the base DiT module. This internal helper walks the wrapper chain and returns the un

Ask for a filesystem path, optionally validating existence. Raises: GoBack: When ``allow_back`` is True and user types 'b'/'back'.

Return the local retry fallback built from the user's caption and metadata. This path is used when the external provider returns an empty, unchan

Build protocol-specific payload and headers.

Calculate PMI-based quality score for generated audio. PMI (Pointwise Mutual Information) removes condition bias: ``score = log P(condition|c

Clamp duration value to GPU memory limit. Args: duration_value: Duration in seconds (can be None or -1 for no limit). llm_handler

Compute task-mode booleans for downstream generation logic.

Generator-aware ``randn`` for per-step forward-diffusion noise. ``generator`` may be ``None`` (fully random), a single ``torch.Generator``, o

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

Scan and return all model directory names starting with 'acestep-5Hz-lm-

Get the effective free VRAM in GB, accounting for PyTorch allocator cache and per-process memory fraction. torch.cuda.mem_get_info() rep

Compute an MD5 hash identifying the currently loaded LoRA adapter weights. Iterates over the handler's LoRA service registry to find adapter weig