Segmentation fault when using set_tokens

[dgtlntv]

Hey, thanks for creating these python bindings!

When I am setting the prompt_tokens with set_tokens like this:

prompt_tokens = []
for i in range(n_segments):
    token_count = self.w.context.full_n_tokens(i)
    for j in range(token_count):
        prompt_tokens.append(self.w.context.full_get_token_id(i, j))
    
self.w.params.set_tokens(prompt_tokens)

I get a segmentation fault if I call

w.context.full(params, audio)

after that.

Am I using the set_tokens wrong? 🤔

[aarnphm]

I can take a look. But it seems correct.

[dgtlntv]

That would be great because I have no idea what is causing the issue here. I'll try to provide some minimal example.

[dgtlntv]

from whispercpp import Whisper
import numpy as np
from scipy.io import wavfile

class WhisperCpp:
    def __init__(self, model) -> None:
        self.w = Whisper.from_pretrained(model)
        self.w.params.with_no_context(False)

    def transcribe(self, audio):
        self.w.context.full(self.w.params, audio)

        n_segments = self.w.context.full_n_segments()

        for i in range(n_segments):
            txt = self.w.context.full_get_segment_text(i)
            print(txt)

        prompt_tokens = []
        for i in range(n_segments):
            token_count = self.w.context.full_n_tokens(i)
            for j in range(token_count):
                prompt_tokens.append(self.w.context.full_get_token_id(i, j))
            
        self.w.params.set_tokens(prompt_tokens)

    def __del__(self):
        self.w.context.free()

def main():
    model = "tiny.en"
    w = WhisperCpp(model)

    audio = wavfile.read("jfk.wav")[1]
    print("audio loaded")

    w.transcribe(audio)
    print("1st one transcribed")
    w.transcribe(audio)
    print("2nd one transcribed")

if __name__ == "__main__":
    main()

With this code I am getting a segmentation fault on the second transcription.

[aarnphm]

Thanks. will try it out when I have some time.