Minimal host object exposing InitServiceMixin for focused unit testing.

Structured object yielded by the trainer, backward-compatible with ``(step, loss, msg)`` tuple unpacking. Extra fields give the UI enough con

Configuration for music generation parameters. Attributes: # Text Inputs caption: A short text prompt describing the desired

5Hz LM Handler for audio code generation

Audio saving and transcoding utility class

Raised when external API calls or response parsing fail.

Raised when encrypted secret read/write operations fail.

Configuration for music generation. Attributes: batch_size: Number of audio samples to generate allow_lm_batch: Whether to al

Fake DiT handler providing initialize_service contract.

Test host exposing the minimum runtime attributes for request helpers.

Minimal host exposing extracted native MLX encode/decode helper methods.

Minimal host implementing dependencies for request helper tests.

Minimal host implementing ``generate_music`` helper dependencies. The host captures helper calls in ``self.calls`` and returns deterministic

ACE-Step Business Logic Handler

Minimal handler stand-in exposing the surface the overlay needs.

Handler stub exposing the attributes used by ``set_use_lora`` / ``set_lora_scale``.

Minimal model facade exposing the surface ``flow_edit`` uses.

Fake LLM handler providing initialize contract.

Host combining both encode mixins with minimal dependencies.

Minimal host implementing methods used by ``IoAudioMixin``.

Temporarily redirect user-local storage into a writable temp directory.

r""" This is the configuration class to store the configuration of a [`AceStepModel`]. It is used to instantiate an AceStep model according to

Stateful wrapper that applies DCW per sampler step. Encapsulates the mode / scaler / wavelet choice so the sampler loop only needs to call ``

Native-keyring or OpenSSL-backed secret store in user-local storage.

Two weight-normalised Conv1d layers (k=7 dilated + k=1) wrapped with Snake1d activations and a residual skip connection.

Snake activation: x + (1/beta) * sin(alpha * x)^2. Parameters ``alpha`` and ``beta`` are stored as 1-D vectors of shape [C] and broadcast ove

Dataset that loads preprocessed tensor files. This is the recommended dataset for training as all tensors are pre-computed: - target_late

Parse request parameters from multiple sources with alias support.

Temporarily offload runtime components and restore them after a task.

Configuration for LoRA training process. Training uses: - Device-aware mixed precision (bf16 on CUDA/XPU, fp16 on MPS, fp32 on CPU) -

Return preconfigured records by job id.

Minimal host object exposing TrainingPresetMixin for focused unit testing.

Builder for creating training datasets from audio files.

Decision payload for whether launcher should apply legacy torch fix. Attributes: should_apply: Whether compatibility fix should be applie

单次请求结果

Host stub for testing chunk decode implementations.

Host stub for testing VaeDecodeMixin orchestration behavior.

Minimal decoder stub returning configured cross-attention outputs.

Handler stub exposing the attributes used by ``load_lora``.

Minimal stand-in for a DiT handler with LoRA attributes.

Capture local-cache set calls for deterministic assertions.

Host object exposing the minimal attributes expected by mixins.

Minimal host implementing ConditioningEmbedMixin dependencies.

Host variant that exposes the real MemoryUtilsMixin._get_vae_dtype implementation. The base _Host overrides ``_get_vae_dtype`` for stability in u

Encoder layer for AceStep model. Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.

Encoder layer for AceStep model. Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.

Encoder layer for AceStep model. Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.

Encoder layer for AceStep model. Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.

Encoder layer for AceStep model. Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.

Encoder layer for AceStep model. Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.

Request payload for auto-labeling dataset samples.

Result of creating a music sample from a natural language query. This is used by the "Simple Mode" / "Inspiration Mode" feature where users

Static configuration for a supported external provider.

Result of formatting user-provided caption and lyrics. This is used by the "Format" feature where users provide caption and lyrics, and t

Typed request payload model for generation jobs. This schema mirrors the historical `api_server` request contract so legacy clients remain co

Result of music generation. Attributes: # Audio Outputs audios: List of audio dictionaries with paths, keys, params s

Raised when the user types 'b' or 'back' at any prompt.

Wrapper module for arbitrary lambda functions. Allows using lambda functions in nn.Sequential by wrapping them in a Module. Useful for s

Wrapper module for arbitrary lambda functions. Allows using lambda functions in nn.Sequential by wrapping them in a Module. Useful f

Wrapper module for arbitrary lambda functions. Allows using lambda functions in nn.Sequential by wrapping them in a Module. Useful for s

Wrapper module for arbitrary lambda functions. Allows using lambda functions in nn.Sequential by wrapping them in a Module. Useful for s

Wrapper module for arbitrary lambda functions. Allows using lambda functions in nn.Sequential by wrapping them in a Module. Useful f

Wrapper module for arbitrary lambda functions. Allows using lambda functions in nn.Sequential by wrapping them in a Module. Useful for s

High-level trainer for ACE-Step LoRA fine-tuning. Uses Lightning Fabric for distributed training and mixed precision. Supports training from

Facade over adapter discovery, registry build, and scaling application. When no adapter names are discoverable, a synthetic ``default`` adapter i

Native MLX implementation of AceStepDiTModel (the diffusion transformer decoder). Mirrors the PyTorch ``AceStepDiTModel`` class exactly:

DataModule for preprocessed tensor files. This is the recommended DataModule for training. It loads pre-computed tensors directly without

Result of music understanding from audio codes. Attributes: # Metadata Fields caption: Generated caption describing the music

Minimal handler stub for LoRA route tests.

Minimal route handler stub used for HTTP integration tests.

Minimal parser stub exposing typed accessors used by request builder.

Minimal request stub supporting async json/form/body accessors.

In-memory store stub for worker helper tests.

Minimal host providing state required by payload assembly tests.

Thread-safe in-memory store for queued/running/completed jobs.

Multi-headed attention module for AceStep model. Implements the attention mechanism from 'Attention Is All You Need' paper, with support

Multi-headed attention module for AceStep model. Implements the attention mechanism from 'Attention Is All You Need' paper, with support

Multi-headed attention module for AceStep model. Implements the attention mechanism from 'Attention Is All You Need' paper, with support

Multi-headed attention module for AceStep model. Implements the attention mechanism from 'Attention Is All You Need' paper, with support

Multi-headed attention module for AceStep model. Implements the attention mechanism from 'Attention Is All You Need' paper, with support

Multi-headed attention module for AceStep model. Implements the attention mechanism from 'Attention Is All You Need' paper, with support

Dataset for ACE-Step LoRA training from raw audio. DEPRECATED: Use PreprocessedTensorDataset instead for better performance. Audio F

Raised when model discovery fails for all candidate endpoints.

GPU configuration based on available memory

Configuration for LoKr (Low-Rank Kronecker) training.

High-level trainer for ACE-Step LoKr fine-tuning.

Minimal write buffer storing the latest non-empty log line.

OpenRouter-compatible model information.

Model pricing information.