ipex-llm/python/llm/example/CPU/PyTorch-Models/Model/openai-whisper

Whisper

In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on general pytorch models, for example Openai Whisper models. For illustration purposes, we utilize the whisper-tiny as a reference Whisper model.

0. Requirements

To run these examples with BigDL-LLM, we have some recommended requirements for your machine, please refer to here for more information.

Example: Recognize Tokens using transcribe() API

In the example recognize.py, we show a basic use case for a Whisper model to conduct transcription using transcribe() API, with BigDL-LLM INT4 optimizations.

1. Install

We suggest using conda to manage environment:

conda create -n llm python=3.9
conda activate llm

pip install bigdl-llm[all] # install bigdl-llm with 'all' option
pip install -U openai-whisper
pip install librosa # required by audio processing

2. Run

python ./recognize.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --repo-id-or-data-path REPO_ID_OR_DATA_PATH --language LANGUAGE

Arguments info:

• --model-name MODEL_NAME: argument defining the model name(tiny, medium, base, etc.) for the Whisper model to be downloaded. It is one of the official model names listed by whisper.available_models(), or path to a model checkpoint containing the model dimensions and the model state_dict. It is default to be 'tiny'.
• --audio-file AUDIO_FILE: argument defining the path of the audio file to be recognized.
• --language LANGUAGE: argument defining language to be transcribed. It is default to be english.

Note

: When loading the model in 4-bit, BigDL-LLM converts linear layers in the model into INT4 format. In theory, a XB model saved in 16-bit will requires approximately 2X GB of memory for loading, and ~0.5X GB memory for further inference.

Please select the appropriate size of the Whisper model based on the capabilities of your machine.

2.1 Client

On client Windows machine, it is recommended to run directly with full utilization of all cores:

python ./recognize.py --audio-file /PATH/TO/AUDIO_FILE

2.2 Server

For optimal performance on server, it is recommended to set several environment variables (refer to here for more information), and run the example with all the physical cores of a single socket.

E.g. on Linux,

# set BigDL-Nano env variables
source bigdl-nano-init

# e.g. for a server with 48 cores per socket
export OMP_NUM_THREADS=48
numactl -C 0-47 -m 0 python ./recognize.py

2.3 Sample Output

whisper-tiny

For audio file(.wav) download from https://www.youtube.com/watch?v=-LIIf7E-qFI, it should be extracted as:

[00:00.000 --> 00:10.000]  I don't know who you are.
[00:10.000 --> 00:15.000]  I don't know what you want.
[00:15.000 --> 00:21.000]  If you're looking for ransom, I can tell you I don't know money, but what I do have.
[00:21.000 --> 00:24.000]  I'm a very particular set of skills.
[00:24.000 --> 00:27.000]  The skills I have acquired are very long career.
[00:27.000 --> 00:31.000]  The skills that make me a nightmare for people like you.
[00:31.000 --> 00:35.000]  If you let my daughter go now, that'll be the end of it.
[00:35.000 --> 00:39.000]  I will not look for you. I will not pursue you.
[00:39.000 --> 00:45.000]  But if you don't, I will look for you. I will find you.
[00:45.000 --> 00:48.000]  And I will kill you.
[00:48.000 --> 00:53.000]  Good luck.
Inference time: xxxx s
-------------------- Output --------------------
 I don't know who you are. I don't know what you want. If you're looking for ransom, I can tell you I don't know money, but what I do have. I'm a very particular set of skills. The skills I have acquired are very long career. The skills that make me a nightmare for people like you. If you let my daughter go now, that'll be the end of it. I will not look for you. I will not pursue you. But if you don't, I will look for you. I will find you. And I will kill you. Good luck.