From 49d636e2953bcfc238087c564866bec6ce7ac71c Mon Sep 17 00:00:00 2001 From: Zhao Changmin Date: Wed, 19 Jul 2023 08:33:20 +0800 Subject: [PATCH] [LLM] whisper model transformer int4 verification and example (#8511) * LLM: transformer api support * va * example * revert * pep8 * pep8 --- .../transformers_int4/whisper/readme.md | 59 +++++++++++++++ .../transformers_int4/whisper/recognized.py | 73 +++++++++++++++++++ .../src/bigdl/llm/transformers/__init__.py | 2 +- .../llm/src/bigdl/llm/transformers/model.py | 4 + 4 files changed, 137 insertions(+), 1 deletion(-) create mode 100644 python/llm/example/transformers/transformers_int4/whisper/readme.md create mode 100644 python/llm/example/transformers/transformers_int4/whisper/recognized.py diff --git a/python/llm/example/transformers/transformers_int4/whisper/readme.md b/python/llm/example/transformers/transformers_int4/whisper/readme.md new file mode 100644 index 00000000..ed273ede --- /dev/null +++ b/python/llm/example/transformers/transformers_int4/whisper/readme.md @@ -0,0 +1,59 @@ +# Whisper + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Whisper models. For illustration purposes, we utilize the [openai/whisper-tiny](https://huggingface.co/openai/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](../README.md#recommended-requirements) for more information. + +## Example: Predict Tokens using `generate()` API +In the example [generate.py](./generate.py), we show a basic use case for a Whisper model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations. +### 1. Install +We suggest using conda to manage environment: +```bash +conda create -n llm python=3.9 +conda activate llm + +pip install bigdl-llm[all] # install bigdl-llm with 'all' option +``` + +### 2. Run +``` +python ./generate.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: +- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Whisper model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'openai/whisper-tiny'`. +- `--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 `'hf-internal-testing/librispeech_asr_dummy'`. +- `--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 *X*B model saved in 16-bit will requires approximately 2*X* GB of memory for loading, and ~0.5*X* 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: +```powershell +python ./generate.py +``` + +#### 2.2 Server +For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration-on-linux) for more information), and run the example with all the physical cores of a single socket. + +E.g. on Linux, +```bash +# 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 ./generate.py +``` + +#### 2.3 Sample Output +#### [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) + +```log +Inference time: 0.23290777206420898 s +-------------------- Output -------------------- +[" Mr. Quilter is the Apostle of the Middle classes and we're glad to welcome his Gospel."] +``` \ No newline at end of file diff --git a/python/llm/example/transformers/transformers_int4/whisper/recognized.py b/python/llm/example/transformers/transformers_int4/whisper/recognized.py new file mode 100644 index 00000000..a367fd9f --- /dev/null +++ b/python/llm/example/transformers/transformers_int4/whisper/recognized.py @@ -0,0 +1,73 @@ +# +# 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. +# + +import torch +import time +import argparse + +from bigdl.llm.transformers import AutoModelForSpeechSeq2Seq +from transformers import WhisperProcessor +from datasets import load_dataset + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Whisper model') + parser.add_argument('--repo-id-or-model-path', type=str, default="openai/whisper-tiny", + help='The huggingface repo id for the whisper model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--repo-id-or-data-path', type=str, + default="hf-internal-testing/librispeech_asr_dummy", + help='The huggingface repo id for the audio dataset to be downloaded' + ', or the path to the huggingface dataset folder') + parser.add_argument('--language', type=str, default="english", + help='language to be transcribed') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + dataset_path = args.repo_id_or_data_path + language = args.language + + # Load model in 4 bit, + # which convert the relevant layers in the model into INT4 format + model = AutoModelForSpeechSeq2Seq.from_pretrained(model_path, + load_in_4bit=True) + model.config.forced_decoder_ids = None + + # Load tokenizer + processor = WhisperProcessor.from_pretrained(model_path) + forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task="transcribe") + + # load dummy dataset and read audio files + ds = load_dataset(dataset_path, "clean", split="validation") + + # Generate predicted tokens + with torch.inference_mode(): + sample = ds[0]["audio"] + + input_features = processor(sample["array"], + sampling_rate=sample["sampling_rate"], + return_tensors="pt").input_features + st = time.time() + # if your selected model is capable of utilizing previous key/value attentions + # to enhance decoding speed, but has `"use_cache": false` in its model config, + # it is important to set `use_cache=True` explicitly in the `generate` function + # to obtain optimal performance with BigDL-LLM INT4 optimizations + predicted_ids = model.generate(input_features, + forced_decoder_ids=forced_decoder_ids) + end = time.time() + output_str = processor.batch_decode(predicted_ids, skip_special_tokens=True) + print(f'Inference time: {end-st} s') + print('-'*20, 'Output', '-'*20) + print(output_str) \ No newline at end of file diff --git a/python/llm/src/bigdl/llm/transformers/__init__.py b/python/llm/src/bigdl/llm/transformers/__init__.py index 1471db04..10ce4fbc 100644 --- a/python/llm/src/bigdl/llm/transformers/__init__.py +++ b/python/llm/src/bigdl/llm/transformers/__init__.py @@ -15,5 +15,5 @@ # from .convert import ggml_convert_quant -from .model import AutoModelForCausalLM, AutoModel, AutoModelForSeq2SeqLM +from .model import AutoModelForCausalLM, AutoModel, AutoModelForSeq2SeqLM, AutoModelForSpeechSeq2Seq from .modelling_bigdl import BigdlNativeForCausalLM diff --git a/python/llm/src/bigdl/llm/transformers/model.py b/python/llm/src/bigdl/llm/transformers/model.py index 85e1015a..2c86c83b 100644 --- a/python/llm/src/bigdl/llm/transformers/model.py +++ b/python/llm/src/bigdl/llm/transformers/model.py @@ -162,5 +162,9 @@ class AutoModel(_BaseAutoModelClass): HF_Model = transformers.AutoModel +class AutoModelForSpeechSeq2Seq(_BaseAutoModelClass): + HF_Model = transformers.AutoModelForSpeechSeq2Seq + + class AutoModelForSeq2SeqLM(_BaseAutoModelClass): HF_Model = transformers.AutoModelForSeq2SeqLM