diff --git a/python/llm/example/GPU/PyTorch-Models/Model/speech-t5/README.md b/python/llm/example/GPU/PyTorch-Models/Model/speech-t5/README.md new file mode 100644 index 00000000..8c78e0d9 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/speech-t5/README.md @@ -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 +
+ +For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series + +```bash +export USE_XETLA=OFF +export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 +``` + +
+ +
+ +For Intel Data Center GPU Max Series + +```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`. +
+ +#### 3.2 Configurations for Windows +
+ +For Intel iGPU + +```cmd +set SYCL_CACHE_PERSISTENT=1 +set BIGDL_LLM_XMX_DISABLED=1 +``` + +
+ +
+ +For Intel Arc™ A300-Series or Pro A60 + +```cmd +set SYCL_CACHE_PERSISTENT=1 +``` + +
+ +
+ +For other Intel dGPU Series + +There is no need to set further environment variables. + +
+ +> 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) diff --git a/python/llm/example/GPU/PyTorch-Models/Model/speech-t5/synthesize_speech.py b/python/llm/example/GPU/PyTorch-Models/Model/speech-t5/synthesize_speech.py new file mode 100644 index 00000000..342d6229 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/speech-t5/synthesize_speech.py @@ -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)