Trait AudioGpu 

pub trait AudioGpu: Send + Sync {
    // Required methods
    fn sample_audio_frame(
        &self,
        embedding: &[f32],
        temperature: f32,
        top_k: usize,
    ) -> [i32; 8];
    fn detokenize_to_spectrum(
        &self,
        cpu_weights: &DetokenizerWeights,
        codes: &[i32],
    ) -> Vec<f32>;
    fn reset_depthformer(&self);
    fn reset_detokenizer(&self);
    fn supports_depthformer(&self) -> bool;

    // Provided methods
    fn sample_audio_frame_async<'a>(
        &'a self,
        embedding: &'a [f32],
        temperature: f32,
        top_k: usize,
    ) -> Pin<Box<dyn Future<Output = Result<[i32; 8]>> + Send + 'a>> { ... }
    fn detokenize_to_spectrum_async<'a>(
        &'a self,
        cpu_weights: &'a DetokenizerWeights,
        codes: &'a [i32],
    ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>>> + Send + 'a>> { ... }
    fn istft_to_pcm(
        &self,
        spectrum: &[f32],
        n_fft: usize,
        hop_length: usize,
    ) -> Vec<f32> { ... }
    fn istft_to_pcm_async<'a>(
        &'a self,
        spectrum: &'a [f32],
        n_fft: usize,
        hop_length: usize,
    ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>>> + Send + 'a>> { ... }
    fn try_acquire_session(&self) -> bool { ... }
    fn release_session(&self) { ... }
}

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

Required Methods

fn sample_audio_frame( &self, embedding: &[f32], temperature: f32, top_k: usize, ) -> [i32; 8]

Sample 8 audio codes from an LLM embedding using the depthformer.

fn detokenize_to_spectrum( &self, cpu_weights: &DetokenizerWeights, codes: &[i32], ) -> Vec<f32>

Convert 8 audio codes to spectrum [n_frames × n_fft_bins × 2].

fn reset_depthformer(&self)

Reset depthformer KV caches (called per audio frame).

fn reset_detokenizer(&self)

Reset detokenizer state (conv buffers + KV caches, called per generation).

fn supports_depthformer(&self) -> bool

Whether AudioGpu::sample_audio_frame is actually implemented here.

Provided Methods

fn sample_audio_frame_async<'a>( &'a self, embedding: &'a [f32], temperature: f32, top_k: usize, ) -> Pin<Box<dyn Future<Output = Result<[i32; 8]>> + Send + 'a>>

Async version of Self::sample_audio_frame for WebGPU / browser wasm.

fn detokenize_to_spectrum_async<'a>( &'a self, cpu_weights: &'a DetokenizerWeights, codes: &'a [i32], ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>>> + Send + 'a>>

Async version of Self::detokenize_to_spectrum.

fn istft_to_pcm( &self, spectrum: &[f32], n_fft: usize, hop_length: usize, ) -> Vec<f32>

Convert the accumulated spectrum [n_frames × n_fft_bins × 2] (log-mag, angle) to PCM via ISTFT. Defaults to the CPU istft_to_pcm; GPU backends override to run the iDFT-matmul + windowed overlap-add on-device.

fn istft_to_pcm_async<'a>( &'a self, spectrum: &'a [f32], n_fft: usize, hop_length: usize, ) -> Pin<Box<dyn Future<Output = Result<Vec<f32>>> + Send + 'a>>

Async version of Self::istft_to_pcm.

fn try_acquire_session(&self) -> bool

Attempt to acquire an exclusive session lease for multi-frame generation. Returns true if acquired, or false if another session is actively generating audio.

fn release_session(&self)

Release the exclusive session lease acquired by Self::try_acquire_session.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors