Builds MossTTSRealtime prompt inputs with text and audio codebooks. This processor focuses on preparing the mixed text/audio token layout expected

Decode audio tokens into waveform chunks with optional crossfade.

Step-wise inference wrapper for MossTTSRealtime. This class mirrors the non-streaming inference logic but exposes a prefill/step/finish API fo

Manage text-to-audio streaming for a single conversation.

Sum-of-embeddings lookup for the MOSS-TTS-Delay multi-channel input.

Wrapper around the Qwen3 backbone running in llama.cpp. Accepts embedding vectors as input (bypassing tok_embd) and returns hidden states (af

Full MOSS-TTS-Delay inference pipeline using llama.cpp backbone. When ``config.low_memory`` is True, GPU-heavy components (backbone, audio en

HF-style MLP adapter equivalent to Megatron's SwiGLU FFN: in: input_size mid: ffn_hidden_size out: output_size Computes:

Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py

Mutable state for one generation sequence (batch_size=1).

Accumulates labelled GPU memory snapshots and prints a summary.

Container returned by :meth:`MossSoundEffectPipeline.__call__`.

Multi-headed attention from 'Attention Is All You Need' paper

r""" This is the configuration class to store the configuration of a [`MossTTSDelayModel`]. It is used to instantiate an MossTTSDelay model ac

r""" This is the configuration class to store the configuration of a [`MossTTSDelayModel`]. It is used to instantiate an MossTTSDelay model ac

Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding). Args: loss (`torch.FloatTe

Teacher-forcing dataset for MossTTSRealtime supervised fine-tuning. All records use the ``conversations`` format. Labels are only set for as

Compute logits from hidden state for all 33 prediction heads (pure NumPy).

Wraps Qwen3 (decoder-only) as a text encoder for Wan audio pipeline. Loads the full Qwen3 model and extracts last-layer hidden states as text

Introduced in SoundStream: An end2end neural audio codec https://arxiv.org/abs/2107.03312