From 6d1ca88aacee7e4d2dca0b0493f31391aeae7560 Mon Sep 17 00:00:00 2001
From: Yina Chen <33650826+cyita@users.noreply.github.com>
Date: Thu, 10 Aug 2023 12:42:14 +0800
Subject: [PATCH] add voice assistant example (#8711)

---
 .../GPU/voiceassistant/README.md              |  47 +++++++
 .../GPU/voiceassistant/generate.py            | 131 ++++++++++++++++++
 2 files changed, 178 insertions(+)
 create mode 100644 python/llm/example/transformers/transformers_int4/GPU/voiceassistant/README.md
 create mode 100644 python/llm/example/transformers/transformers_int4/GPU/voiceassistant/generate.py

diff --git a/python/llm/example/transformers/transformers_int4/GPU/voiceassistant/README.md b/python/llm/example/transformers/transformers_int4/GPU/voiceassistant/README.md
new file mode 100644
index 00000000..d797ea1b
--- /dev/null
+++ b/python/llm/example/transformers/transformers_int4/GPU/voiceassistant/README.md
@@ -0,0 +1,47 @@
+# Voice Assistant
+In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Whisper and Llama2 models on any Intel® Arc™ A-Series Graphics. For illustration purposes, we utilize the following models: 
+- [openai/whisper-small](https://huggingface.co/openai/whisper-small) and [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) as reference whisper models.
+- [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) and [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf) as reference Llama2 models.
+
+## 0. Requirements
+To run these examples with BigDL-LLM on Intel® Arc™ A-Series Graphics, 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 conduct transcription using `generate()` API, then use the recoginzed text as the input for Llama2 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel® Arc™ A-Series Graphics.
+### 1. Install
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9
+conda activate llm
+# below command will install intel_extension_for_pytorch==2.0.110+xpu as default
+# you can install specific ipex/torch version for your need
+pip install bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+pip install librosa soundfile datasets
+pip install accelerate
+pip install SpeechRecognition sentencepiece colorama
+```
+### 2. Configures OneAPI environment variables
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+
+### 3. Run
+
+For optimal performance on Arc, it is recommended to set several environment variables.
+
+```bash
+export USE_XETLA=OFF
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+```
+
+```
+python ./generate.py --llama2-repo-id-or-model-path REPO_ID_OR_MODEL_PATH --whisper-repo-id-or-model-path REPO_ID_OR_MODEL_PATH --n-predict N_PREDICT
+```
+
+Arguments info:
+- `--llama2-repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Llama2 model (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'meta-llama/Llama-2-7b-chat-hf'`.
+- `--whisper-repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Whisper model (e.g. `openai/whisper-small` and `openai/whisper-medium`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'openai/whisper-small'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+#### Sample Output
+Should be tested on a linux machine with microphone.
diff --git a/python/llm/example/transformers/transformers_int4/GPU/voiceassistant/generate.py b/python/llm/example/transformers/transformers_int4/GPU/voiceassistant/generate.py
new file mode 100644
index 00000000..568d2fea
--- /dev/null
+++ b/python/llm/example/transformers/transformers_int4/GPU/voiceassistant/generate.py
@@ -0,0 +1,131 @@
+#
+# 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 os
+import torch
+import time
+import argparse
+import numpy as np
+
+from bigdl.llm.transformers import AutoModelForCausalLM
+from bigdl.llm.transformers import AutoModelForSpeechSeq2Seq
+from transformers import LlamaTokenizer
+import intel_extension_for_pytorch as ipex
+from transformers import WhisperProcessor
+from transformers import TextStreamer
+from colorama import Fore
+import speech_recognition as sr
+from datasets import load_dataset
+
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://huggingface.co/georgesung/llama2_7b_chat_uncensored#prompt-style
+DEFAULT_SYSTEM_PROMPT = """\
+"""
+
+def get_prompt(message: str, chat_history: list[tuple[str, str]],
+               system_prompt: str) -> str:
+    texts = [f'<s>[INST] <<SYS>>\n{system_prompt}\n<</SYS>>\n\n']
+    # The first user input is _not_ stripped
+    do_strip = False
+    for user_input, response in chat_history:
+        user_input = user_input.strip() if do_strip else user_input
+        do_strip = True
+        texts.append(f'{user_input} [/INST] {response.strip()} </s><s>[INST] ')
+    message = message.strip() if do_strip else message
+    texts.append(f'{message} [/INST]')
+    return ''.join(texts)
+
+def get_input_features(r):
+    with sr.Microphone(device_index=1, sample_rate=16000) as source:
+        print("Calibrating...")
+        r.adjust_for_ambient_noise(source, duration=5)
+
+        print(Fore.YELLOW + "Listening now..." + Fore.RESET)
+        try:
+            audio = r.listen(source, timeout=5, phrase_time_limit=30)
+            # refer to https://github.com/openai/whisper/blob/main/whisper/audio.py#L63
+            frame_data = np.frombuffer(audio.frame_data, np.int16).flatten().astype(np.float32) / 32768.0
+            input_features = processor(frame_data, sampling_rate=audio.sample_rate, return_tensors="pt").input_features
+            input_features = input_features.half().contiguous().to('xpu')
+            print("Recognizing...")
+        except Exception as e:
+            unrecognized_speech_text = (
+                f"Sorry, I didn't catch that. Exception was: \n {e}"
+            )
+            print(unrecognized_speech_text)
+    
+    return input_features
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
+    parser.add_argument('--llama2-repo-id-or-model-path', type=str, default="meta-llama/Llama-2-7b-chat-hf",
+                        help='The huggingface repo id for the Llama2 (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--whisper-repo-id-or-model-path', type=str, default="openai/whisper-small",
+                        help='The huggingface repo id for the Whisper (e.g. `openai/whisper-small` and `openai/whisper-medium`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--n-predict', type=int, default=32,
+                        help='Max tokens to predict')
+
+    args = parser.parse_args()
+    whisper_model_path = args.whisper_repo_id_or_model_path
+    llama_model_path = args.llama2_repo_id_or_model_path
+
+    dataset_path = "hf-internal-testing/librispeech_asr_dummy"
+    # Load dummy dataset and read audio files
+    ds = load_dataset(dataset_path, "clean", split="validation")
+
+    print("Converting and loading models...")
+    processor = WhisperProcessor.from_pretrained(whisper_model_path)
+
+    # generate token ids
+    whisper =  AutoModelForSpeechSeq2Seq.from_pretrained(whisper_model_path, load_in_4bit=True, optimize_model=False)
+    whisper.config.forced_decoder_ids = None
+    whisper = whisper.half().to('xpu')
+    
+    llama_model = AutoModelForCausalLM.from_pretrained(llama_model_path, load_in_4bit=True, trust_remote_code=True, optimize_model=False)
+    llama_model = llama_model.half().to('xpu')
+    tokenizer = LlamaTokenizer.from_pretrained(llama_model_path)
+
+    r = sr.Recognizer()
+
+    with torch.inference_mode():
+        # warm up
+        sample = ds[2]["audio"]
+        input_features = processor(sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt").input_features
+        input_features = input_features.half().contiguous().to('xpu')
+        torch.xpu.synchronize()
+        predicted_ids = whisper.generate(input_features)
+        torch.xpu.synchronize()
+        output_str = processor.batch_decode(predicted_ids, skip_special_tokens=True)
+        output_str = output_str[0]
+        input_ids = tokenizer.encode(output_str, return_tensors="pt").to('xpu')
+        output = llama_model.generate(input_ids, do_sample=False, max_new_tokens=32)
+        output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+        torch.xpu.synchronize()
+        
+        while 1:
+            input_features = get_input_features(r)
+            predicted_ids = whisper.generate(input_features)
+            output_str = processor.batch_decode(predicted_ids, skip_special_tokens=True)
+            output_str = output_str[0]
+            print("\n" + Fore.GREEN + "Whisper : " + Fore.RESET + "\n" + output_str)
+            print("\n" + Fore.BLUE + "BigDL-LLM: " + Fore.RESET)
+            prompt = get_prompt(output_str, [], system_prompt=DEFAULT_SYSTEM_PROMPT)
+            input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
+            streamer = TextStreamer(tokenizer, skip_special_tokens=True, skip_prompt=True)
+            _ = llama_model.generate(input_ids, streamer=streamer, do_sample=False, max_new_tokens=args.n_predict)