ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

LLM: add speech T5 GPU example (#10090)

Browse source 
* add speech t5 example

* fix

* fix
This commit is contained in:
Jin Qiao 2024-02-07 10:50:02 +08:00 committed by GitHub
parent 2f4c754759
commit d3d2ee1b63
2 changed files with 222 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

121
python/llm/example/GPU/PyTorch-Models/Model/speech-t5/README.md Normal file
View file

@ -0,0 +1,121 @@
# SpeechT5
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate SpeechT5 models. For illustration purposes, we utilize the [microsoft/speecht5_tts](https://huggingface.co/microsoft/speecht5_tts) as reference SpeechT5 models.
## Requirements
To run these examples with BigDL-LLM, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-requirements) for more information.
## Example: Synthesize speech with the given input text
In the example [synthesize_speech.py](./synthesize_speech.py), we show a basic use case for SpeechT5 model to synthesize speech based on the given text, with BigDL-LLM INT4 optimizations.
### 1. Install
#### 1.1 Installation on Linux
We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
After installing conda, create a Python environment for BigDL-LLM:
```bash
conda create -n llm python=3.9 # recommend to use Python 3.9
conda activate llm
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
pip install datasets soundfile # additional package required for SpeechT5 to conduct generation
```
#### 1.2 Installation on Windows
We suggest using conda to manage environment:
```bash
conda create -n llm python=3.9 libuv
conda activate llm
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
pip install datasets soundfile # additional package required for SpeechT5 to conduct generation
```
### 2. Configures OneAPI environment variables
#### 2.1 Configurations for Linux
```bash
source /opt/intel/oneapi/setvars.sh
```
#### 2.2 Configurations for Windows
```cmd
call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
```
> Note: Please make sure you are using **CMD** (**Anaconda Prompt** if using conda) to run the command as PowerShell is not supported.
### 3. Runtime Configurations
For optimal performance, it is recommended to set several environment variables. Please check out the suggestions based on your device.
#### 3.1 Configurations for Linux
<details>
<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
```bash
export USE_XETLA=OFF
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
```
</details>
<details>
<summary>For Intel Data Center GPU Max Series</summary>
```bash
export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export ENABLE_SDP_FUSION=1
```
> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
</details>
#### 3.2 Configurations for Windows
<details>
<summary>For Intel iGPU</summary>
```cmd
set SYCL_CACHE_PERSISTENT=1
set BIGDL_LLM_XMX_DISABLED=1
```
</details>
<details>
<summary>For Intel Arc™ A300-Series or Pro A60</summary>
```cmd
set SYCL_CACHE_PERSISTENT=1
```
</details>
<details>
<summary>For other Intel dGPU Series</summary>
There is no need to set further environment variables.
</details>
> Note: For the first time that each model runs on Intel iGPU/Intel Arc™ A300-Series or Pro A60, it may take several minutes to compile.
### 4. Running examples
```bash
python ./synthesize_speech.py --text 'Artificial intelligence refers to the development of computer systems that can perform tasks that typically require human intelligence.'
```
In the example, several arguments can be passed to satisfy your requirements:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the SpeechT5 model (e.g `microsoft/speecht5_tts`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/speecht5_tts'`.
- `--repo-id-or-vocoder-path REPO_ID_OR_VOCODER_PATH`: argument defining the huggingface repo id for the SpeechT5 vocoder (e.g `microsoft/speecht5_hifigan`, which generates audio from a spectrogram) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/speecht5_hifigan'`.
- `--repo-id-or-data-path REPO_ID_OR_DATA_PATH`: argument defining the huggingface repo id for the audio dataset (e.g. `Matthijs/cmu-arctic-xvectors`, which decides voice characteristics) to be downloaded, or the path to the huggingface dataset folder. It is default to be `'Matthijs/cmu-arctic-xvectors'`.
- `--text TEXT`: argument defining the text to synthesize speech. It is default to be `"Artificial intelligence refers to the development of computer systems that can perform tasks that typically require human intelligence."`.
#### 4.1 Sample Output
#### [microsoft/speecht5_tts](https://huggingface.co/microsoft/speecht5_tts)
Text: Artificial intelligence refers to the development of computer systems that can perform tasks that typically require human intelligence.
[Click here to hear sample output.](https://llm-assets.readthedocs.io/en/latest/_downloads/f0bebfbe8c350b71fe565a82192c079b/speech-t5-example-output.wav)

101
python/llm/example/GPU/PyTorch-Models/Model/speech-t5/synthesize_speech.py Normal file
View file

@ -0,0 +1,101 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Some parts of this file is adapted from
# https://huggingface.co/microsoft/speecht5_tts
#
# MIT License
#
# Copyright (c) Microsoft Corporation.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE
import torch
import time
import argparse
from bigdl.llm import optimize_model
from transformers import SpeechT5Processor, SpeechT5HifiGan, SpeechT5ForTextToSpeech
from datasets import load_dataset
import soundfile as sf
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Synthesize speech with the given input text using SpeechT5 model')
parser.add_argument('--repo-id-or-model-path', type=str, default='microsoft/speecht5_tts',
help='The huggingface repo id for the SpeechT5 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--repo-id-or-vocoder-path', type=str, default='microsoft/speecht5_hifigan',
help='The huggingface repo id for the vocoder model (which generates audio from a spectrogram) to be downloaded'
', or the path to the huggingface checkpoint folder.')
parser.add_argument('--repo-id-or-data-path', type=str, default="Matthijs/cmu-arctic-xvectors",
help='The huggingface repo id for the audio dataset (which decides voice characteristics) to be downloaded'
', or the path to the huggingface dataset folder')
parser.add_argument('--text', type=str, default="Artificial intelligence refers to the development of computer systems that can perform tasks that typically require human intelligence.",
help='Text to synthesize speech')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
vocoder_path = args.repo_id_or_vocoder_path
dataset_path = args.repo_id_or_data_path
text = args.text
processor = SpeechT5Processor.from_pretrained(model_path)
model = SpeechT5ForTextToSpeech.from_pretrained(model_path)
vocoder = SpeechT5HifiGan.from_pretrained(vocoder_path)
# With only one line to enable BigDL-LLM optimization on model
# Skip optimizing these two modules to get higher audio quality
# When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the optimize_model function.
# This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
model = optimize_model(model, modules_to_not_convert=["speech_decoder_postnet.feat_out",
"speech_decoder_postnet.prob_out"])
model = model.to('xpu')
vocoder = vocoder.to('xpu')
inputs = processor(text=text, return_tensors="pt").to('xpu')
# load xvector containing speaker's voice characteristics from a dataset
embeddings_dataset = load_dataset(dataset_path, split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0).to('xpu')
with torch.inference_mode():
# ipex model needs a warmup, then inference time can be accurate
speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
st = time.time()
speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
torch.xpu.synchronize()
end = time.time()
print(f"Inference time: {end-st} s")
sf.write("bigdl_llm_speech_t5_out.wav", speech.to('cpu').numpy(), samplerate=16000)