mirror of
https://github.com/RVC-Boss/GPT-SoVITS.git
synced 2025-12-16 09:16:59 +08:00
d3b8f7e09e
This commit migrates the project from using NVIDIA CUDA to Intel XPU for GPU acceleration, based on the PyTorch 2.9 release. Key changes include: - Replaced `torch.cuda` with `torch.xpu` for device checks, memory management, and distributed training. - Updated device strings from "cuda" to "xpu" across the codebase. - Switched the distributed training backend from "nccl" to "ccl" for Intel GPUs. - Disabled custom CUDA kernels in the `BigVGAN` module by setting `use_cuda_kernel=False`. - Updated `requirements.txt` to include `torch==2.9` and `intel-extension-for-pytorch`. - Modified CI/CD pipelines and build scripts to remove CUDA dependencies and build for an XPU target.
88 lines
2.6 KiB
Python
88 lines
2.6 KiB
Python
# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
|
|
# LICENSE is in incl_licenses directory.
|
|
|
|
from __future__ import absolute_import, division, print_function, unicode_literals
|
|
|
|
import os
|
|
import argparse
|
|
import json
|
|
import torch
|
|
import librosa
|
|
from utils import load_checkpoint
|
|
from meldataset import get_mel_spectrogram
|
|
from scipy.io.wavfile import write
|
|
from env import AttrDict
|
|
from meldataset import MAX_WAV_VALUE
|
|
from bigvgan import BigVGAN as Generator
|
|
|
|
h = None
|
|
device = None
|
|
torch.backends.cudnn.benchmark = False
|
|
|
|
|
|
def inference(a, h):
|
|
generator = Generator(h, use_cuda_kernel=a.use_cuda_kernel).to(device)
|
|
|
|
state_dict_g = load_checkpoint(a.checkpoint_file, device)
|
|
generator.load_state_dict(state_dict_g["generator"])
|
|
|
|
filelist = os.listdir(a.input_wavs_dir)
|
|
|
|
os.makedirs(a.output_dir, exist_ok=True)
|
|
|
|
generator.eval()
|
|
generator.remove_weight_norm()
|
|
with torch.no_grad():
|
|
for i, filname in enumerate(filelist):
|
|
# Load the ground truth audio and resample if necessary
|
|
wav, sr = librosa.load(os.path.join(a.input_wavs_dir, filname), sr=h.sampling_rate, mono=True)
|
|
wav = torch.FloatTensor(wav).to(device)
|
|
# Compute mel spectrogram from the ground truth audio
|
|
x = get_mel_spectrogram(wav.unsqueeze(0), generator.h)
|
|
|
|
y_g_hat = generator(x)
|
|
|
|
audio = y_g_hat.squeeze()
|
|
audio = audio * MAX_WAV_VALUE
|
|
audio = audio.cpu().numpy().astype("int16")
|
|
|
|
output_file = os.path.join(a.output_dir, os.path.splitext(filname)[0] + "_generated.wav")
|
|
write(output_file, h.sampling_rate, audio)
|
|
print(output_file)
|
|
|
|
|
|
def main():
|
|
print("Initializing Inference Process..")
|
|
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--input_wavs_dir", default="test_files")
|
|
parser.add_argument("--output_dir", default="generated_files")
|
|
parser.add_argument("--checkpoint_file", required=True)
|
|
# --use_cuda_kernel argument is removed to disable custom CUDA kernels.
|
|
# parser.add_argument("--use_cuda_kernel", action="store_true", default=False)
|
|
|
|
a = parser.parse_args()
|
|
a.use_cuda_kernel = False
|
|
|
|
config_file = os.path.join(os.path.split(a.checkpoint_file)[0], "config.json")
|
|
with open(config_file) as f:
|
|
data = f.read()
|
|
|
|
global h
|
|
json_config = json.loads(data)
|
|
h = AttrDict(json_config)
|
|
|
|
torch.manual_seed(h.seed)
|
|
global device
|
|
if torch.xpu.is_available():
|
|
torch.xpu.manual_seed(h.seed)
|
|
device = torch.device("xpu")
|
|
else:
|
|
device = torch.device("cpu")
|
|
|
|
inference(a, h)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|