Struct WhisperRunner 

pub struct WhisperRunner {
    pub device: Device,
    pub activation_dtype: DType,
    /* private fields */
}

Fields

device: Device

activation_dtype: DType

Implementations

impl WhisperRunner

pub fn builder() -> WhisperRunnerBuilder

pub fn config(&self) -> &WhisperConfig

pub fn decode_buckets_compiled(&self) -> usize

Number of bucketed decode graphs compiled so far (bench / tuning).

pub fn mel_frames(&self) -> usize

pub fn enc_seq(&self) -> usize

pub fn decode_device(&self) -> Device

Device that runs bucketed decode graphs (may differ from Self::device on Metal/MLX).

pub fn stage_device(&self) -> Device

Device that runs encoder, cross, prefill, and decode graphs.

pub fn uses_gpu_kv(&self) -> bool

pub fn encode_mel(&mut self, mel: &MelSpectrogram) -> Result<Vec<f32>, Error>

pub fn encode_pcm(&mut self, samples: &[f32]) -> Result<Vec<f32>, Error>

pub fn encode_wav(&mut self, path: &Path) -> Result<Vec<f32>, Error>

pub fn prefill_prompt( &mut self, cross: &LayerKvCache, prompt_tokens: &[u32], batch: usize, ) -> Result<(Vec<f32>, LayerKvCache), Error>

pub fn swap_decode_cache(&mut self, other: &mut WhisperRunner)

Exchange bucketed decode compile caches (precision: share CPU-compiled graphs).

pub fn decode_one_step( &mut self, cross: &LayerKvCache, token: u32, cache: &mut LayerKvCache, ) -> Result<Vec<f32>, Error>

Single greedy decode step (for cross-backend parity checks).

pub fn encode_mel_batch( &mut self, mels: &[MelSpectrogram], ) -> Result<Vec<f32>, Error>

pub fn bench_greedy_pipeline( &mut self, pcm: &[f32], decode_steps: usize, warmup: usize, ) -> Result<(WhisperBenchReport, String), Error>

Per-stage greedy-decode timings (after optional warmup). Includes last prefill logits for parity checks.

pub fn bench_greedy_from_encoder( &mut self, enc: &[f32], decode_steps: usize, ) -> Result<(WhisperBenchReport, String), Error>

Greedy decode benchmark from a fixed encoder output (cross-backend precision: share CPU enc).

pub fn bench_greedy_from_cross( &mut self, cross: &LayerKvCache, decode_steps: usize, ) -> Result<(WhisperBenchReport, String), Error>

Greedy decode from a fixed cross-attention cache (share CPU cross for precision).

pub fn bench_greedy_decode_from_state( &mut self, cross: &LayerKvCache, prompt: &[u32], prefill_logits: Vec<f32>, cache: LayerKvCache, decode_steps: usize, ) -> Result<(WhisperBenchReport, String), Error>

Greedy decode from CPU prefill logits + KV (cross-backend decode parity).

pub fn cross_cache_batch( &mut self, enc: &[f32], batch: usize, ) -> Result<LayerKvCache, Error>

pub fn transcribe_greedy(&mut self, pcm: &[f32]) -> Result<String, Error>

pub fn transcribe_beam(&mut self, pcm: &[f32]) -> Result<String, Error>

pub fn transcribe_with_vad(&mut self, pcm: &[f32]) -> Result<String, Error>

pub fn transcribe_regions_batched( &mut self, pcm: &[f32], regions: &[SpeechSegment], beam_size: usize, ) -> Result<Vec<String>, Error>

pub fn build_prompt(&self) -> Result<Vec<u32>, Error>