Gadersd/whisper-burn

WGPU issues in Macbook M1

DavidGOrtega opened this issue · 5 comments

DavidGOrtega commented

As a TVM user Im very excited of this project because of the use of burn and its access to WGPU native. Personally speaking is the way to go.
However my tests are very discouraging. WGPU seems to be performing worse than CPU

WGPU

cargo run --release --bin transcribe --features wgpu-backend  medium  audio16k.wav transcription.txt

     Running `target/release/transcribe medium audio16k.wav en transcription.txt`
Loading waveform...
Loading model...
Depth: 0
...
Chunk 0:  Hello, I am the whisper machine learning model. If you see this as text, then I am working properly.

infer took: 49665 ms

CPU

cargo run --release --bin transcribe medium audio16k.wav en transcription.txt

    Running `target/release/transcribe medium audio16k.wav en transcription.txt`
Loading waveform...
Loading model...
Depth: 0
...
Chunk 0:  Hello, I am the whisper machine learning model. If you see this as text, then I am working properly.

infer took: 19517 ms
Transcription finished.

the code was slightly modified:

fn main() {
    cfg_if::cfg_if! {
        if #[cfg(feature = "wgpu-backend")] {
            type Backend = WgpuBackend<AutoGraphicsApi, f32, i32>;
            let device = WgpuDevice::BestAvailable;
        } else if #[cfg(feature = "torch-backend")] {
            type Backend = TchBackend<f32>;
            let device = TchDevice::Cpu;
        }
    }


... 

let start_time = Instant::now();
    let (text, tokens) = match waveform_to_text(&whisper, &bpe, lang, waveform, sample_rate) {
        Ok((text, tokens)) => (text, tokens),
        Err(e) => {
            eprintln!("Error during transcription: {}", e);
            process::exit(1);
        }
    };
    let end_time = Instant::now();
    let elapsed_time_ms = end_time.duration_since(start_time).as_millis();

    println!("infer took: {} ms", elapsed_time_ms);

Same 3X for tiny CPU vs tiny WGPU

Might not be optimised for my machine? It's not working maybe?

DavidGOrtega commented

Mps performance is also quite bad, similar to WGPU

type Backend = TchBackend<f32>;
let device = TchDevice::Mps;
antimora commented

We need to review/profile if the stft (featurizer) is slow. I believe it was implemented manually and might be slow compared to specialized libraries.

  • 👍1
DavidGOrtega commented

thanks for the reply @antimora Im going to profile and see

Gadersd commented

I haven't yet prioritized optimization. Caching should speed up the inference significantly. I don't think the burn-wgpu backend has been significantly optimized yet. You might want to check with its maintainers.

  • 👍1