Module audio_decoder 

Audio decoder for LFM2.5-Audio: vocoder GGUF loading + weight structures.

Three sub-models loaded from the vocoder GGUF:

1. DecoderModel — samples 8 audio codes per frame from LLM embedding
2. DepthformerModel — small 6-layer attention-only transformer (backbone of DecoderModel)
3. Detokenizer — codes → spectrogram → PCM (not loaded here yet — Phase 5)

Structs

AudioDecoderWeights

All weights for the audio decoder (loaded from vocoder GGUF).

CodebookWeights

Per-codebook embedding layer weights.

DecoderConfig

Configuration for the decoder model (8-codebook sampling).

DepthformerConfig

Configuration for the depthformer (small transformer inside the decoder).

DepthformerLayerWeights

Per-layer weights for one depthformer layer. Uses MmapWeight for GEMV to match ggml’s Q4_0 computation path.

DepthformerState

Runtime state for the depthformer (reset per audio frame).

DetokLayerWeights

Weights for one detokenizer LFM2 layer.

DetokenizerConfig

Config for the detokenizer’s LFM2 backbone.

DetokenizerState

Runtime state for the detokenizer’s LFM2 backbone.

DetokenizerWeights

All detokenizer weights (loaded from vocoder GGUF).

IstftStreamer

Stateful streaming inverse STFT (overlap-add) processor.

Traits

AudioGpu

GPU-accelerated audio backend. Implementations provide Metal or WGPU dispatch for the depthformer (code sampling) and detokenizer (spectrum).

Functions

build_gpu_audio_decoder

Try to construct a GPU audio decoder backend for the given vocoder GGUF file.

build_hann

Periodic Hann window of length n_fft: w[n] = 0.5·(1 - cos(2π·n/n_fft)).

build_idft_basis

Build the real inverse-DFT basis for the GPU ISTFT.

depthformer_forward

Run one forward pass through the depthformer (1 token, n_embd input → n_embd output).

depthformer_forward_inplace

In-place depthformer forward pass directly reading state.depthformer_in.

depthformer_forward_into

In-place depthformer forward pass that avoids allocating a returning Vec.

detok_embed_codes

Embed 8 audio codes into a single vector for the detokenizer. Looks up each code with per-codebook offset, averages across codebooks.

detokenize_to_spectrum

Run the detokenizer: codes → spectrogram frames (before ISTFT).

embed_audio_token

Convert 8 audio codes back into an embedding for feeding to the LLM.

istft_to_pcm

Convert spectrogram frames to PCM audio samples via ISTFT.

sample_audio_frame

Sample one audio frame (8 codes) from an LLM embedding.

upsample

Linear interpolation upsample: 1 token → n_up tokens.