LLM: Fix discards in optimize_model with non-hf models and add openai whisper example (#8877)

* openai-whisper

python/llm/example/pytorch-model/openai-whisper/readme.md

@@ -0,0 +1,74 @@
+# 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](https://github.com/openai/whisper/blob/main/model-card.md) 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: Recognize Tokens using `transcribe()` API
+In the example [recognize.py](./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:
+```bash
+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 *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 ./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](../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 ./recognize.py
+```
+
+#### 2.3 Sample Output
+#### [whisper-tiny](https://github.com/openai/whisper/blob/main/model-card.md)
+
+For audio file(.wav) download from https://www.youtube.com/watch?v=-LIIf7E-qFI, it should be extracted as:
+```log
+[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.
+```

python/llm/example/pytorch-model/openai-whisper/recognize.py

@@ -0,0 +1,58 @@
+#
+# 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 whisper
+import time
+import librosa
+import argparse
+from bigdl.llm import optimize_model
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Recognize Tokens using `transcribe()` API for Openai Whisper model')
+    parser.add_argument('--model-name', type=str, default="tiny",
+                        help="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.")
+    parser.add_argument('--audio-file', type=str, required=True,
+                        help='The path of the audio file to be recognized.')
+    parser.add_argument('--language', type=str, default="English",
+                        help='language to be transcribed')
+    args = parser.parse_args()
+
+    # Load the input audio
+    y, sr = librosa.load(args.audio_file)
+
+    # Downsample the audio to 16kHz
+    target_sr = 16000
+    audio = librosa.resample(y,
+                            orig_sr=sr,
+                            target_sr=target_sr)
+
+    # Load whisper model under pytorch framework
+    model = whisper.load_model(args.model_name)
+
+    # With only one line to enable bigdl optimize on a pytorch model
+    model = optimize_model(model)
+
+    st = time.time()
+    result = model.transcribe(audio, verbose=True, language=args.language)
+    end = time.time()
+    print(f'Inference time: {end-st} s')
+
+    print('-'*20, 'Output', '-'*20)
+    print(result["text"])

python/llm/src/bigdl/llm/transformers/convert.py

@@ -41,6 +41,7 @@ from accelerate import init_empty_weights
 import warnings
 import transformers
 import importlib
+from .utils import logger
 
 
 def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
@@ -129,6 +130,14 @@ def convert_forward(m, target_m, new_forward):
 def optimize(model):
     from packaging import version
     from bigdl.llm.transformers.models.llama import llama_attention_forward_4_31
+    from transformers.modeling_utils import PreTrainedModel
+
+    # All huggingface format models are inherited from `PreTrainedModel`
+    if not isinstance(model, PreTrainedModel):
+        logger.info("It is not supported that optimizing a model isn't belong to"
+                    "huggingface transformers with `optimize_model=True` for now.")
+        return model
+
     trans_version = transformers.__version__
     if version.parse(trans_version) >= version.parse("4.31.0"):
         convert_forward(

python/llm/src/bigdl/llm/transformers/model.py

@@ -23,10 +23,7 @@ from bigdl.llm.ggml.quantize import ggml_tensor_qtype
 from bigdl.llm.utils.common import invalidInputError
 import torch
 import copy
-import logging
-
-logging.basicConfig(format='%(asctime)s - %(levelname)s - %(message)s', level=logging.INFO)
-logger = logging.getLogger(__name__)
+from .utils import logger
 
 
 def save_low_bit(self, *args, **kwargs):

python/llm/src/bigdl/llm/transformers/utils.py

@@ -45,6 +45,10 @@ from ..utils.common import invalidInputError
 from typing import Union
 import torch
 from torch import nn
+import logging
+
+logging.basicConfig(format='%(asctime)s - %(levelname)s - %(message)s', level=logging.INFO)
+logger = logging.getLogger(__name__)
 
 
 WEIGHTS_NAME = "pytorch_model.bin"