引言

docker pull quay.io/ascend/vllm-ascend:v0.11.0rc2-openeuler

docker images | grep v0.11.0rc2-openeuler

gitcode download Ascend-SACT/Kokoro
cd Kokoro
unzip kokoro_img_v1.tar.gz.zip
docker load -i kokoro_img_v1.tar.gz

docker images | grep kokoro_image

docker run -itd -u root \
--net=host \
--privileged=true \
--device=/dev/davinci0 \
--device=/dev/davinci1 \
--device=/dev/davinci2 \
--device=/dev/davinci3 \
--device=/dev/davinci4 \
--device=/dev/davinci5 \
--device=/dev/davinci6 \
--device=/dev/davinci7 \
--device=/dev/davinci_manager \
--device=/dev/devmm_svm \
--device=/dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /root:/root \
-p 8001:8001 \
--shm-size 1024g \
--name vllm-ascend-wlh2 \
quay.io/ascend/vllm-ascend:v0.11.0rc2-openeuler

docker exec -it -u root vllm-ascend-wlh2 bash

pip install -q kokoro
pip install -q soundfile
pip install -q "misaki[zh]>=0.8.2"

pip install en_core_web_sm-3.8.0-py3-none-any.whl

# This file is hardcoded to transparently reproduce HEARME_zh.wav
# Therefore it may NOT generalize gracefully to other texts
# Refer to Usage in README.md for more general usage patterns

# pip install kokoro>=0.8.1 "misaki[zh]>=0.8.1"
from kokoro import KModel, KPipeline
from pathlib import Path
import numpy as np
import soundfile as sf
import torch
import torch_npu
from torch_npu.contrib import transfer_to_npu
import tqdm
import os

MODLE_DIR = '/home/w00634320/Kokoro-82M-v1.1-zh'
MODLE_PATH = os.path.join(MODLE_DIR, 'kokoro-v1_1-zh.pth')
CONFIG_PATH = os.path.join(MODLE_DIR, 'config.json')
REPO_ID = "hexgrad/Kokoro-82M-v1.1-zh"
VOICE_PATH = os.path.join(MODLE_DIR, "voices/zf_001.pt")
SAMPLE_RATE = 24000
EN_VOICE_PATH = os.path.join(MODLE_DIR, "voices/af_maple.pt")
# How much silence to insert between paragraphs: 5000 is about 0.2 seconds
N_ZEROS = 5000

# Whether to join sentences in paragraphs 1 and 3
JOIN_SENTENCES = True

VOICE = 'zf_001' if True else 'zm_010'

device = 'npu'

texts = [(
"Kokoro 是一系列体积虽小但功能强大的 TTS 模型。",
), (
"该模型是经过短期训练的结果，从专业数据集中添加了100名中文使用者。",
"中文数据由专业数据集公司「龙猫数据」免费且无偿地提供给我们。感谢你们让这个模型成为可能。",
), (
"另外，一些众包合成英语数据也进入了训练组合：",
"1小时的 Maple，美国女性。",
"1小时的 Sol，另一位美国女性。",
"和1小时的 Vale，一位年长的英国女性。",
), (
"由于该模型删除了许多声音，因此它并不是对其前身的严格升级，但它提前发布以收集有关新声音和标记化的反馈。",
"除了中文数据集和3小时的英语之外，其余数据都留在本次训练中。",
"目标是推动模型系列的发展，并最终恢复一些被遗留的声音。",
), (
"美国版权局目前的指导表明，合成数据通常不符合版权保护的资格。",
"由于这些合成数据是众包的，因此模型训练师不受任何服务条款的约束。",
"该 Apache 许可模式也符合 OpenAI 所宣称的广泛传播 AI 优势的使命。",
"如果您愿意帮助进一步完成这一使命，请考虑为此贡献许可的音频数据。",
)]

if JOIN_SENTENCES:
    for i in (1, 3):
        texts[i] = [''.join(texts[i])]

# HACK: Mitigate rushing caused by lack of training data beyond ~100 tokens
# Simple piecewise linear fn that decreases speed as len_ps increases
def speed_callable(len_ps):
    speed = 0.8
    if len_ps <= 83:
        speed = 1
    elif len_ps < 183:
        speed = 1 - (len_ps - 83) / 500
    return speed * 1.1

en_pipeline = KPipeline(lang_code='a', repo_id=REPO_ID, model=False)
def en_callable(text):
    if text == 'Kokoro':
        return 'kˈOkəɹO'
    elif text == 'Sol':
        return 'sˈOl'
    return next(en_pipeline(text)).phonemes
model = KModel(repo_id=REPO_ID, config=CONFIG_PATH, model=MODLE_PATH).to(device).eval()
zh_pipeline = KPipeline(lang_code='z', repo_id=REPO_ID, model=model, en_callable=en_callable)

path = Path(__file__).parent

wavs = []
for paragraph in tqdm.tqdm(texts):
    for i, sentence in enumerate(paragraph):
        generator = zh_pipeline(sentence, voice=VOICE_PATH, speed=speed_callable)
        f = path / f'zh{len(wavs):02}.wav'
        result = next(generator)
        wav = result.audio
        sf.write(f, wav, SAMPLE_RATE)
        if i == 0 and wavs and N_ZEROS > 0:
            wav = np.concatenate([np.zeros(N_ZEROS), wav])
        wavs.append(wav)

sf.write(path / f'HEARME_{VOICE}.wav', np.concatenate(wavs), SAMPLE_RATE)

     def transform(self, input_data):
         forward_transform = torch.stft(
             input_data,
             self.filter_length, self.hop_length, self.win_length, window=self.window.to(input_data.device),
             return_complex=True)
         real_part = forward_transform.real
         imag_part = forward_transform.imag
         return torch.sqrt(real_part**2 + imag_part**2), torch.angle(forward_transform)

source /usr/local/Ascend/ascend-toolkit/set_env.sh
python test_infer.py

# This file is hardcoded to transparently reproduce HEARME_zh.wav
# Therefore it may NOT generalize gracefully to other texts
# Refer to Usage in README.md for more general usage patterns

# pip install kokoro>=0.8.1 "misaki[zh]>=0.8.1"
from kokoro import KModel, KPipeline
from pathlib import Path
import numpy as np
import soundfile as sf
import torch
import torch_npu
from torch_npu.contrib import transfer_to_npu
import tqdm
import os
import time
import threading

MODLE_DIR = '/home/xxxx/Kokoro-82M-v1.1-zh'
MODLE_PATH = os.path.join(MODLE_DIR, 'kokoro-v1_1-zh.pth')
CONFIG_PATH = os.path.join(MODLE_DIR, 'config.json')
REPO_ID = "hexgrad/Kokoro-82M-v1.1-zh"
VOICE_PATH = os.path.join(MODLE_DIR, "voices/zf_001.pt")
SAMPLE_RATE = 24000

# How much silence to insert between paragraphs: 5000 is about 0.2 seconds
N_ZEROS = 5000

# Whether to join sentences in paragraphs 1 and 3
JOIN_SENTENCES = True

VOICE = 'zf_001' if True else 'zm_010'

device = 'npu'

texts = [
    "春天，是一年四季中最富有生机的季节。天气渐渐变暖，阳光变得柔和，大地披上了翠绿的新装。树木抽出嫩芽，花朵竞相开放，五彩斑斓，像是给世界增添了无数的欢笑。田野里，农民伯伯开始忙碌地播种，期待着秋天的丰收。公园中，孩子们奔跑嬉戏，老人在树下散步聊天，一切都显得那么和谐美好。",
    "夏日的阳光像一把金色的琴弦，弹奏出热烈而欢快的旋律。蝉鸣是夏日最忠实的乐手，在树梢上不知疲倦地唱着。午后，阳光炙烤着大地，连空气都仿佛在微微颤动。但夏天也有它的温柔时刻，比如傍晚的凉风轻拂过脸颊，带来一丝清凉，或是夜晚微星光亮中，一家人围坐在院子里吃着西瓜，谈笑风生。",
    "秋风拂过，带来了凉爽与成熟的味道。金黄的稻田在阳光下闪闪发光，仿佛大地披上了一件华丽的外衣。树叶从枝头缓缓飘落，像一只只翩翩起舞的蝴蝶，点缀着大地的画卷。秋天是收获的季节，果园里果实累累，苹果、柿子、葡萄，香甜的气息弥漫在空气中。人们忙着采摘，脸上洋溢着丰收的喜悦。",
    "冬天悄然而至，寒风裹挟着雪花，为大地披上一层银装。清晨推开窗，白茫茫的世界仿佛被按下暂停键，宁静而纯净。树枝披上冰晶，屋檐垂下冰凌，阳光洒在雪地上，反射出耀眼的光芒。孩子们在雪地里奔跑、打雪仗、堆雪人，欢声笑语驱散了寒冷。"
]


# HACK: Mitigate rushing caused by lack of training data beyond ~100 tokens
# Simple piecewise linear fn that decreases speed as len_ps increases
def speed_callable(len_ps):
    speed = 0.8
    if len_ps <= 83:
        speed = 1
    elif len_ps < 183:
        speed = 1 - (len_ps - 83) / 500
    return speed * 1.1

model = KModel(repo_id=REPO_ID, config=CONFIG_PATH, model=MODLE_PATH).to(device).eval()
zh_pipeline = KPipeline(lang_code='z', repo_id=REPO_ID, model=model)

path = Path(__file__).parent

def test_warmup():
    sentence= "中华人民共和国万岁"
    generator = zh_pipeline(sentence, voice=VOICE_PATH, speed=speed_callable)
    result = next(generator)
    end_time = time.time()
    wav = result.audio

def test_inf(sentence, num):
    f = path / f'test_{num}.wav'
    start_time = time.time()
    generator = zh_pipeline(sentence, voice=VOICE_PATH, speed=speed_callable)

    result = next(generator)
    end_time = time.time()
    wav = result.audio
    speech_len = len(wav) / SAMPLE_RATE
    print('yield speech len {}, rtf {}'.format(speech_len, (end_time - start_time) / speech_len))
    sf.write(f, wav, SAMPLE_RATE)

test_warmup()
ths = []
for i in range(len(texts)):
    t = threading.Thread(target=test_inf, name='LoopThread' + str(i), args=(texts[i], i))
    ths.append(t)
    t.start()

for t in ths:
    t.join()