[LLM] whisper model transformer int4 verification and example (#8511)

* LLM: transformer api support

* va

* example

* revert

* pep8

* pep8

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."]
+```

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)

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

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