Files
Lumotia/crates/transcription/src/whisper_rs_backend.rs
jars 4cb954ece4
Some checks failed
check / cargo check (macos-latest) (push) Has been cancelled
check / cargo check (ubuntu-22.04) (push) Has been cancelled
check / cargo check (windows-latest) (push) Has been cancelled
check / svelte build + lint (push) Has been cancelled
perf: route whisper + llm n_threads through physical-core helper
Both inference call sites previously called `num_cpus::get()` (logical
thread count). Established whisper.cpp / llama.cpp guidance is that
SMT siblings contend for shared FPU resources during heavy F16/F32
matmul, so going past physical core count anti-scales. Empirical
sweep on Whisper Tiny / 11s JFK clip / Ryzen 5 4650U (6c12t):

  n_threads | xc_time | RTF    | speedup_vs_1
  ----------|---------|--------|-------------
          1 |   0.33s |  0.030 |   1.00x
          2 |   0.33s |  0.030 |   1.00x
          4 |   0.30s |  0.028 |   1.09x
          6 |   0.32s |  0.029 |   1.04x
          8 |   0.31s |  0.028 |   1.06x
         12 |   0.32s |  0.029 |   1.03x

Tiny doesn't scale (work dominated by overhead) but the larger
Whisper variants and Qwen LLMs do. Sources: whisper.cpp #200, #1033,
#1252, #403; llama.cpp #3167, #572.

Changes:
- crates/core/src/constants.rs:
  * MIN_INFERENCE_THREADS lowered 4 → 2 (research-derived floor)
  * MAX_INFERENCE_THREADS = 8 added (research-derived ceiling)
  * inference_thread_count() rewritten:
      - reads MAGNOTIA_INFERENCE_THREADS env var (override)
      - num_cpus::get_physical() with available_parallelism fallback
      - clamped to [MIN_INFERENCE_THREADS, MAX_INFERENCE_THREADS]
- crates/core/Cargo.toml: + num_cpus = "1"
- crates/transcription/src/whisper_rs_backend.rs: call site uses helper.
- crates/llm/src/lib.rs: call site uses helper.
- crates/transcription/Cargo.toml: drop num_cpus from [features] +
  [dependencies] (production no longer needs it). Move to
  [dev-dependencies] for tests/thread_sweep.rs only.
- crates/llm/Cargo.toml: drop num_cpus = "1" (no longer used directly).

Per-machine maps (after this patch):
  Ryzen 5 4650U  6c12t → 6
  big-iron 12c24t       → 8 (clamp; users can override)
  cheap 2c2t laptop     → 2
  1c container/VM       → 2

Adds crates/transcription/tests/thread_sweep.rs — env-gated like
jfk_bench, prints the table above against any model + WAV. Useful for
re-baselining on new hardware or when tuning the clamp values.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-09 09:26:23 +01:00

126 lines
4.2 KiB
Rust

//! Direct whisper-rs backend. Owns a WhisperContext; each call builds a
//! fresh WhisperState (state can be reused, but fresh-per-call is simpler
//! and matches the transcribe-rs call style we are replacing).
//!
//! Exists because transcribe-rs does not expose set_initial_prompt; this
//! wrapper is the only path that can pipe per-capture vocabulary context
//! into Whisper.
use std::path::Path;
use whisper_rs::{FullParams, SamplingStrategy, WhisperContext, WhisperContextParameters};
use magnotia_core::constants::inference_thread_count;
use magnotia_core::error::{MagnotiaError, Result};
use magnotia_core::types::{Segment, TranscriptionOptions};
use crate::transcriber::{Transcriber, TranscriberCapabilities};
#[derive(Debug, thiserror::Error)]
pub enum WhisperBackendError {
#[error("whisper-rs load failed: {0}")]
Load(String),
#[error("whisper-rs state creation failed: {0}")]
State(String),
#[error("whisper-rs transcribe failed: {0}")]
Transcribe(String),
}
pub struct WhisperRsBackend {
ctx: WhisperContext,
}
impl WhisperRsBackend {
pub fn load(model_path: &Path) -> std::result::Result<Self, WhisperBackendError> {
let ctx = WhisperContext::new_with_params(model_path, WhisperContextParameters::default())
.map_err(|e| WhisperBackendError::Load(e.to_string()))?;
Ok(Self { ctx })
}
}
impl Transcriber for WhisperRsBackend {
fn capabilities(&self) -> TranscriberCapabilities {
TranscriberCapabilities {
sample_rate: magnotia_core::constants::WHISPER_SAMPLE_RATE,
channels: 1,
supports_initial_prompt: true,
}
}
/// Synchronously transcribe 16 kHz mono f32 PCM.
///
/// `options.initial_prompt` is piped directly to whisper-rs — this
/// is the only backend path that honours it; `SpeechModelAdapter`
/// discards it (Parakeet has no equivalent).
fn transcribe_sync(
&mut self,
samples: &[f32],
options: &TranscriptionOptions,
) -> Result<Vec<Segment>> {
tracing::info!(
language = ?options.language,
has_initial_prompt = options.initial_prompt.as_deref().map(|p| !p.is_empty()).unwrap_or(false),
"WhisperRsBackend::transcribe_sync entering"
);
let mut state = self.ctx.create_state().map_err(|e| {
MagnotiaError::TranscriptionFailed(WhisperBackendError::State(e.to_string()).to_string())
})?;
let mut params = FullParams::new(SamplingStrategy::Greedy { best_of: 1 });
if let Some(lang) = options.language.as_deref() {
if !lang.is_empty() {
params.set_language(Some(lang));
}
}
if let Some(prompt) = options.initial_prompt.as_deref() {
if !prompt.is_empty() {
params.set_initial_prompt(prompt);
}
}
params.set_n_threads(inference_thread_count() as i32);
params.set_print_special(false);
params.set_print_progress(false);
params.set_print_realtime(false);
state.full(params, samples).map_err(|e| {
MagnotiaError::TranscriptionFailed(
WhisperBackendError::Transcribe(e.to_string()).to_string(),
)
})?;
let n = state.full_n_segments();
let mut out = Vec::with_capacity(n.max(0) as usize);
for i in 0..n {
let Some(seg) = state.get_segment(i) else {
continue;
};
let text = seg
.to_str()
.map_err(|e| {
MagnotiaError::TranscriptionFailed(
WhisperBackendError::Transcribe(e.to_string()).to_string(),
)
})?
.to_string();
// whisper-rs timestamps are centiseconds (10ms units). Convert to seconds (f64).
let start = seg.start_timestamp() as f64 * 0.01;
let end = seg.end_timestamp() as f64 * 0.01;
out.push(Segment { start, end, text });
}
Ok(out)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn backend_error_displays() {
let e = WhisperBackendError::Load("oops".into());
assert!(e.to_string().contains("oops"));
}
}