intel arc a770 support

[SlyRebula]

Do you guys have any plans to add intel arc gpu support ?

[hbredin]

pyannote is based on PyTorch for GPU inference: any hardware that PyTorch support is therefore supported.

[SlyRebula]

there is a fork of pytorch by intel that has a support for my gpu i can run whisper on it but when i try to run pyannote i got this error and no output, it gets stuck.

python ./Main.py
C:\Users\SlyRebula\miniconda3\envs\whisperr\lib\site-packages\pyannote\audio\core\io.py:43: UserWarning: torchaudio._backend.set_audio_backend has been deprecated. With dispatcher enabled, this function is no-op. You can remove the function call.
  torchaudio.set_audio_backend("soundfile")
FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

FFT WARNING: INPUT_STRIDES and OUTPUT_STRIDES are deprecated: please use FWD_STRIDES and BWD_STRIDES, instead.

The Source Code:

from pyannote.audio import Pipeline
import intel_extension_for_pytorch
pipeline = Pipeline.from_pretrained(
    "pyannote/speaker-diarization-3.1",
    use_auth_token="itsasecret!")

# send pipeline to GPU (when available)
import torch
pipeline.to(torch.device("xpu"))

# apply pretrained pipeline
diarization = pipeline("test.mp3")

# print the result
for turn, _, speaker in diarization.itertracks(yield_label=True):
    print(f"start={turn.start:.1f}s stop={turn.end:.1f}s speaker_{speaker}")