ipex-llm/python/llm/example/GPU/HF-Transformers-AutoModels/Model/distil-whisper/README.md

Distil-Whisper

In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Distil-Whisper models on Intel GPUs. For illustration purposes, we utilize the distil-whisper/distil-large-v2 as a reference Distil-Whisper model.

0. Requirements

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

Example: Recognize Tokens using generate() API

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

1. Install

We suggest using conda to manage the Python environment. For more information about conda installation, please refer to here.

After installing conda, create a Python environment for BigDL-LLM:

conda create -n llm 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 librosa # required by audio processing

2. Configures OneAPI environment variables

source /opt/intel/oneapi/setvars.sh

3. Run

For optimal performance on Arc, it is recommended to set several environment variables.

export USE_XETLA=OFF
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1

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 --chunk-length CHUNK_LENGTH --batch-size BATCH_SIZE

Arguments info:

• --repo-id-or-model-path REPO_ID_OR_MODEL_PATH: argument defining the huggingface repo id for the Distil-Whisper model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be 'distil-whisper/distil-large-v2'.
• --repo-id-or-data-path REPO_ID_OR_DATA_PATH: argument defining the huggingface repo id for the audio dataset to be downloaded, or the path to the huggingface dataset folder. It is default to be 'distil-whisper/librispeech_long'.
• --language LANGUAGE: argument defining language to be transcribed. It is default to be english.
• --chunk-length CHUNK_LENGTH: argument defining the maximum number of chuncks of sampling_rate samples used to trim and pad longer or shorter audio sequences. For audio recordings less than 30 seconds, it can be set to 0 for better performance. It is default to be 15.
• --batch-size BATCH_SIZE: argument defining the batch_size of pipeline inference, it usually equals of length of the audio divided by chunk-length. It is default to be 16.

Sample Output

Short-Form Transcription

Model: distil-whisper/distil-large-v2

Command line:

python ./recognize.py --repo-id-or-data-path 'hf-internal-testing/librispeech_asr_dummy' --chunk-length 0

Output:

Inference time: xxxx s
-------------------- Output --------------------
[' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.']

Long-Form Transcription

Model: distil-whisper/distil-large-v2

Command line:

python ./recognize.py --repo-id-or-data-path 'distil-whisper/librispeech_long' --chunk-length 15

Output:

inference time is xxxx s
 Mr Quilter is the Apostle of the Middle classes, and we are glad to welcome his Gospel. Nor is Mr Quilter's manner less interesting than his matter. He tells us that at this festive season of the year, with Christmas and roast beef looming before us, similes drawn from eating and its results occur most readily to the mind. He has grave doubts whether Sir Frederick Leighton's work is really Greek after all, and can discover in it but little of rocky Ithaca. Linel's pictures are a sort of upguards and Adam paintings, and Mason's exquisite itels are as national as a Jingo poem. Mr Birkett Foster's landscapes smile at one much in the same way that Mr. Karker used to flash his teeth, and Mr. John Collier gives his sitter a cheerful slap on the back before he says, like a shampoo or a Turkish bath, next man.