ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

LLM: add chatglm3 examples (#9305)

Browse source
This commit is contained in:
Jin Qiao 2023-11-01 09:50:05 +08:00 committed by GitHub
parent 06447a3ef6
commit c44c6dc43a
17 changed files with 899 additions and 80 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
README.md
View file

@ -135,6 +135,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
| LLaMA 2 | [link1](python/llm/example/CPU/Native-Models), [link2](python/llm/example/CPU/HF-Transformers-AutoModels/Model/llama2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/llama2) |
| ChatGLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/chatglm) | |
| ChatGLM2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/chatglm2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/chatglm2) |
| ChatGLM3 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/chatglm3) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/chatglm3) |
| Mistral | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/mistral) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/mistral) |
| Falcon | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/falcon) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/falcon) |
| MPT | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/mpt) | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/mpt) |

1
python/llm/README.md
View file

@ -42,6 +42,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
| LLaMA 2 | [link1](example/CPU/Native-Models), [link2](example/CPU/HF-Transformers-AutoModels/Model/llama2) | [link](example/GPU/HF-Transformers-AutoModels/Model/llama2) |
| ChatGLM | [link](example/CPU/HF-Transformers-AutoModels/Model/chatglm) | |
| ChatGLM2 | [link](example/CPU/HF-Transformers-AutoModels/Model/chatglm2) | [link](example/GPU/HF-Transformers-AutoModels/Model/chatglm2) |
| ChatGLM3 | [link](example/CPU/HF-Transformers-AutoModels/Model/chatglm3) | [link](example/GPU/HF-Transformers-AutoModels/Model/chatglm3) |
| Mistral | [link](example/CPU/HF-Transformers-AutoModels/Model/mistral) | [link](example/GPU/HF-Transformers-AutoModels/Model/mistral) |
| Falcon | [link](example/CPU/HF-Transformers-AutoModels/Model/falcon) | [link](example/GPU/HF-Transformers-AutoModels/Model/falcon) |
| MPT | [link](example/CPU/HF-Transformers-AutoModels/Model/mpt) | [link](example/CPU/HF-Transformers-AutoModels/Model/mpt) |

27
python/llm/example/CPU/HF-Transformers-AutoModels/Model/README.md
View file

@ -1,33 +1,6 @@
# BigDL-LLM Transformers INT4 Optimization for Large Language Model
You can use BigDL-LLM to run any Huggingface Transformer models with INT4 optimizations on either servers or laptops. This directory contains example scripts to help you quickly get started using BigDL-LLM to run some popular open-source models in the community. Each model has its own dedicated folder, where you can find detailed instructions on how to install and run it.
# Verified models
| Model | Example |
|-----------|----------------------------------------------------------|
| LLaMA | [link](vicuna) |
| LLaMA 2 | [link](llama2) |
| MPT | [link](mpt) |
| Falcon | [link](falcon) |
| ChatGLM | [link](chatglm) |
| ChatGLM2 | [link](chatglm2) |
| MOSS | [link](moss) |
| Baichuan | [link](baichuan) |
| Baichuan2 | [link](baichuan2) |
| Dolly-v1 | [link](dolly_v1) |
| Dolly-v2 | [link](dolly_v2) |
| RedPajama | [link](redpajama) |
| Phoenix | [link](phoenix) |
| StarCoder | [link](starcoder) |
| InternLM | [link](internlm) |
| Whisper | [link](whisper) |
| Qwen | [link](qwen) |
| Aquila | [link](aquila) |
| Replit | [link](replit) |
| Mistral | [link](mistral) |
| Flan-t5 | [link](flan-t5) |
| Phi-1_5 | [link](phi-1_5) |
| Qwen-VL | [link](qwen-vl) |
## Recommended Requirements
To run the examples, we recommend using Intel® Xeon® processors (server), or >= 12th Gen Intel® Core™ processor (client).

129
python/llm/example/CPU/HF-Transformers-AutoModels/Model/chatglm3/README.md Normal file
View file

@ -0,0 +1,129 @@
# ChatGLM3
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on ChatGLM3 models. For illustration purposes, we utilize the [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) as a reference ChatGLM3 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 1: Predict Tokens using `generate()` API
In the example [generate.py](./generate.py), we show a basic use case for a ChatGLM3 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 --pre --upgrade bigdl-llm[all] # install bigdl-llm with 'all' option
```
### 2. Run
```
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
```
Arguments info:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM3 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
> **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 ChatGLM3 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
#### [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
<|user|>
AI是什么？
<|assistant|>
-------------------- Output --------------------
[gMASK]sop <|user|>
AI是什么？
<|assistant|> AI是人工智能（Artificial Intelligence）的缩写，指的是通过计算机程序和算法模拟人类智能的技术。AI可以帮助我们解决各种问题，例如语音
```
```log
Inference time: xxxx s
-------------------- Prompt --------------------
<|user|>
What is AI?
<|assistant|>
-------------------- Output --------------------
[gMASK]sop <|user|>
What is AI?
<|assistant|>
AI stands for Artificial Intelligence. It refers to the development of computer systems that can perform tasks that would normally require human intelligence, such as recognizing speech or making
```
## Example 2: Stream Chat using `stream_chat()` API
In the example [streamchat.py](./streamchat.py), we show a basic use case for a ChatGLM3 model to stream chat, 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 --pre --upgrade bigdl-llm[all] # install bigdl-llm with 'all' option
```
### 2. Run
**Stream Chat using `stream_chat()` API**:
```
python ./streamchat.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --question QUESTION
```
**Chat using `chat()` API**:
```
python ./streamchat.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --question QUESTION --disable-stream
```
Arguments info:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM3 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--question QUESTION`: argument defining the question to ask. It is default to be `"晚上睡不着应该怎么办"`.
- `--disable-stream`: argument defining whether to stream chat. If include `--disable-stream` when running the script, the stream chat is disabled and `chat()` API is used.
> **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 ChatGLM3 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
$env:PYTHONUNBUFFERED=1 # ensure stdout and stderr streams are sent straight to terminal without being first buffered
python ./streamchat.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
export PYTHONUNBUFFERED=1 # ensure stdout and stderr streams are sent straight to terminal without being first buffered
numactl -C 0-47 -m 0 python ./streamchat.py
```

69
python/llm/example/CPU/HF-Transformers-AutoModels/Model/chatglm3/generate.py Normal file
View file

@ -0,0 +1,69 @@
#
# 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
import numpy as np
from bigdl.llm.transformers import AutoModel
from transformers import AutoTokenizer
# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM3 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--prompt', type=str, default="AI是什么？",
help='Prompt to infer')
parser.add_argument('--n-predict', type=int, default=32,
help='Max tokens to predict')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
# Load model in 4 bit,
# which convert the relevant layers in the model into INT4 format
model = AutoModel.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt")
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
output = model.generate(input_ids,
max_new_tokens=args.n_predict)
end = time.time()
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
print(f'Inference time: {end-st} s')
print('-'*20, 'Prompt', '-'*20)
print(prompt)
print('-'*20, 'Output', '-'*20)
print(output_str)

62
python/llm/example/CPU/HF-Transformers-AutoModels/Model/chatglm3/streamchat.py Normal file
View file

@ -0,0 +1,62 @@
#
# 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
import numpy as np
from bigdl.llm.transformers import AutoModel
from transformers import AutoTokenizer
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Stream Chat for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM3 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--question', type=str, default="晚上睡不着应该怎么办",
help='Qustion you want to ask')
parser.add_argument('--disable-stream', action="store_true",
help='Disable stream chat')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
disable_stream = args.disable_stream
# Load model in 4 bit,
# which convert the relevant layers in the model into INT4 format
model = AutoModel.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
with torch.inference_mode():
if disable_stream:
# Chat
response, history = model.chat(tokenizer, args.question, history=[])
print('-'*20, 'Chat Output', '-'*20)
print(response)
else:
# Stream chat
response_ = ""
print('-'*20, 'Stream Chat Output', '-'*20)
for response, history in model.stream_chat(tokenizer, args.question, history=[]):
print(response.replace(response_, ""), end="")
response_ = response

14
python/llm/example/CPU/PyTorch-Models/Model/README.md
View file

@ -1,20 +1,6 @@
# BigDL-LLM INT4 Optimization for Large Language Model
You can use `optimize_model` API to accelerate general PyTorch models on Intel servers and PCs. This directory contains example scripts to help you quickly get started using BigDL-LLM to run some popular open-source models in the community. Each model has its own dedicated folder, where you can find detailed instructions on how to install and run it.
# Verified models
| Model | Example |
|----------------|----------------------------------------------------------|
| LLaMA 2 | [link](llama2) |
| ChatGLM | [link](chatglm) |
| Openai Whisper | [link](openai-whisper) |
| BERT | [link](bert) |
| Bark | [link](bark) |
| Mistral | [link](mistral) |
| Flan-t5 | [link](flan-t5) |
| Phi-1_5 | [link](phi-1_5) |
| Qwen-VL | [link](qwen-vl) |
| LLaVA | [link](llava) |
## Recommended Requirements
To run the examples, we recommend using Intel® Xeon® processors (server), or >= 12th Gen Intel® Core™ processor (client).

59
python/llm/example/CPU/PyTorch-Models/Model/chatglm3/README.md Normal file
View file

@ -0,0 +1,59 @@
# ChatGLM3
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate ChatGLM3 models. For illustration purposes, we utilize the [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) as a reference ChatGLM3 model.
## 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 ChatGLM3 model to predict the next N tokens using `generate()` API, 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](https://docs.conda.io/en/latest/miniconda.html#).
After installing conda, create a Python environment for BigDL-LLM:
```bash
conda create -n llm python=3.9 # recommend to use Python 3.9
conda activate llm
pip install --pre --upgrade bigdl-llm[all] # install the latest bigdl-llm nightly build with 'all' option
```
### 2. Run
After setting up the Python environment, you could run the example by following steps.
#### 2.1 Client
On client Windows machines, it is recommended to run directly with full utilization of all cores:
```powershell
python ./generate.py --prompt 'AI是什么？'
```
More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
#### 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 --prompt 'AI是什么？'
```
More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
#### 2.3 Arguments Info
In the example, several arguments can be passed to satisfy your requirements:
- `--repo-id-or-model-path`: str, argument defining the huggingface repo id for the ChatGLM model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`.
#### 2.4 Sample Output
#### [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b)
```log
Inference time: xxxx s
-------------------- Output --------------------
[gMASK]sop <|user|>
AI是什么？
<|assistant|> AI是人工智能（Artificial Intelligence）的缩写，指的是通过计算机程序和算法模拟人类智能的技术。AI可以帮助我们解决各种问题，例如语音
```

61
python/llm/example/CPU/PyTorch-Models/Model/chatglm3/generate.py Normal file
View file

@ -0,0 +1,61 @@
#
# 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 transformers import AutoModel, AutoTokenizer
from bigdl.llm import optimize_model
# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--prompt', type=str, default="AI是什么？",
help='Prompt to infer')
parser.add_argument('--n-predict', type=int, default=32,
help='Max tokens to predict')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
# Load model
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
# With only one line to enable BigDL-LLM optimization on model
model = optimize_model(model)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt")
st = time.time()
output = model.generate(input_ids,
max_new_tokens=args.n_predict)
end = time.time()
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
print(f'Inference time: {end-st} s')
print('-'*20, 'Output', '-'*20)
print(output_str)

25
python/llm/example/GPU/HF-Transformers-AutoModels/Model/README.md
View file

@ -1,31 +1,6 @@
# BigDL-LLM Transformers INT4 Optimization for Large Language Model on Intel GPUs
You can use BigDL-LLM to run almost every Huggingface Transformer models with INT4 optimizations on your laptops with Intel GPUs. This directory contains example scripts to help you quickly get started using BigDL-LLM to run some popular open-source models in the community. Each model has its own dedicated folder, where you can find detailed instructions on how to install and run it.
## Verified models
| Model | Example |
|----------------|----------------------------------------------------------|
| Aquila | [link](aquila) |
| Baichuan | [link](baichuan) |
| Baichuan2 | [link](baichuan2) |
| ChatGLM2 | [link](chatglm2) |
| Chinese Llama2 | [link](chinese-llama2) |
| Dolly v1 | [link](dolly-v1) |
| Dolly v2 | [link](dolly-v2) |
| Falcon | [link](falcon) |
| GPT-J | [link](gpt-j) |
| InternLM | [link](internlm) |
| LLaMA 2 | [link](llama2) |
| Mistral | [link](mistral) |
| MPT | [link](mpt) |
| Qwen | [link](qwen) |
| StarCoder | [link](starcoder) |
| Vicuna | [link](vicuna) |
| Whisper | [link](whisper) |
| Replit | [link](replit) |
| Flan-t5 | [link](flan-t5) |
## Verified Hardware Platforms
- Intel Arc™ A-Series Graphics

109
python/llm/example/GPU/HF-Transformers-AutoModels/Model/chatglm3/README.md Normal file
View file

@ -0,0 +1,109 @@
# ChatGLM3
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on ChatGLM3 models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) as a reference ChatGLM3 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](../README.md#recommended-requirements) for more information.
## Example 1: Predict Tokens using `generate()` API
In the example [generate.py](./generate.py), we show a basic use case for a ChatGLM3 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs.
### 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 --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
```
### 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 --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
```
Arguments info:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM3 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
#### Sample Output
#### [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
<|user|>
AI是什么？
<|assistant|>
-------------------- Output --------------------
[gMASK]sop <|user|>
AI是什么？
<|assistant|> AI是人工智能(Artificial Intelligence)的缩写,指通过计算机程序或机器学习算法来模拟、延伸或扩展人类智能的技术。AI旨在
```
```log
Inference time: xxxx s
-------------------- Prompt --------------------
<|user|>
What is AI?
<|assistant|>
-------------------- Output --------------------
[gMASK]sop <|user|>
What is AI?
<|assistant|>
AI stands for Artificial Intelligence. It refers to the development of computer systems or machines that can perform tasks that would normally require human intelligence, such as recognizing patterns
```
## Example 2: Stream Chat using `stream_chat()` API
In the example [streamchat.py](./streamchat.py), we show a basic use case for a ChatGLM3 model to stream chat, 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
# 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 --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
```
### 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
```
**Stream Chat using `stream_chat()` API**:
```
python ./streamchat.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --question QUESTION
```
**Chat using `chat()` API**:
```
python ./streamchat.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --question QUESTION --disable-stream
```
Arguments info:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM3 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--question QUESTION`: argument defining the question to ask. It is default to be `"晚上睡不着应该怎么办"`.
- `--disable-stream`: argument defining whether to stream chat. If include `--disable-stream` when running the script, the stream chat is disabled and `chat()` API is used.

79
python/llm/example/GPU/HF-Transformers-AutoModels/Model/chatglm3/generate.py Normal file
View file

@ -0,0 +1,79 @@
#
# 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 intel_extension_for_pytorch as ipex
import time
import argparse
import numpy as np
from bigdl.llm.transformers import AutoModel
from transformers import AutoTokenizer
# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM3 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--prompt', type=str, default="AI是什么？",
help='Prompt to infer')
parser.add_argument('--n-predict', type=int, default=32,
help='Max tokens to predict')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
# Load model in 4 bit,
# which convert the relevant layers in the model into INT4 format
model = AutoModel.from_pretrained(model_path,
load_in_4bit=True,
optimize_model=True,
trust_remote_code=True,
use_cache=True)
model = model.to('xpu')
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
# ipex model needs a warmup, then inference time can be accurate
output = model.generate(input_ids,
max_new_tokens=args.n_predict)
# start inference
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
output = model.generate(input_ids,
max_new_tokens=args.n_predict)
torch.xpu.synchronize()
end = time.time()
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
print(f'Inference time: {end-st} s')
print('-'*20, 'Prompt', '-'*20)
print(prompt)
print('-'*20, 'Output', '-'*20)
print(output_str)

72
python/llm/example/GPU/HF-Transformers-AutoModels/Model/chatglm3/streamchat.py Normal file
View file

@ -0,0 +1,72 @@
#
# 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 intel_extension_for_pytorch as ipex
import time
import argparse
import numpy as np
from bigdl.llm.transformers import AutoModel
from transformers import AutoTokenizer
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Stream Chat for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM3 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--question', type=str, default="晚上睡不着应该怎么办",
help='Qustion you want to ask')
parser.add_argument('--disable-stream', action="store_true",
help='Disable stream chat')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
disable_stream = args.disable_stream
# Load model in 4 bit,
# which convert the relevant layers in the model into INT4 format
model = AutoModel.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True,
optimize_model=True)
model.to('xpu')
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
with torch.inference_mode():
prompt = args.question
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
# ipex model needs a warmup, then inference time can be accurate
output = model.generate(input_ids,
max_new_tokens=32)
# start inference
if disable_stream:
# Chat
response, history = model.chat(tokenizer, args.question, history=[])
print('-'*20, 'Chat Output', '-'*20)
print(response)
else:
# Stream chat
response_ = ""
print('-'*20, 'Stream Chat Output', '-'*20)
for response, history in model.stream_chat(tokenizer, args.question, history=[]):
print(response.replace(response_, ""), end="")
response_ = response

14
python/llm/example/GPU/PyTorch-Models/Model/README.md
View file

@ -1,20 +1,6 @@
# BigDL-LLM INT4 Optimization for Large Language Model on Intel GPUs
You can use `optimize_model` API to accelerate general PyTorch models on Intel GPUs. This directory contains example scripts to help you quickly get started using BigDL-LLM to run some popular open-source models in the community. Each model has its own dedicated folder, where you can find detailed instructions on how to install and run it.
## Verified models
| Model | Example |
|----------------|----------------------------------------------------------|
| Mistral | [link](mistral) |
| LLaMA 2 | [link](llama2) |
| ChatGLM2 | [link](chatglm2) |
| Baichuan | [link](baichuan) |
| Baichuan2 | [link](baichuan2) |
| Replit | [link](replit) |
| StarCoder | [link](starcoder) |
| Dolly v1 | [link](dolly-v1) |
| Dolly v2 | [link](dolly-v2) |
| Flan-t5 | [link](flan-t5) |
## Verified Hardware Platforms
- Intel Arc™ A-Series Graphics

108
python/llm/example/GPU/PyTorch-Models/Model/chatglm3/README.md Normal file
View file

@ -0,0 +1,108 @@
# ChatGLM3
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate ChatGLM3 models. For illustration purposes, we utilize the [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) as reference ChatGLM3 models.
## Requirements
To run these examples with BigDL-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-requirements) for more information.
## Example 1: Predict Tokens using `generate()` API
In the example [generate.py](./generate.py), we show a basic use case for a ChatGLM3 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs.
### 1. Install
We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
After installing conda, create a Python environment for BigDL-LLM:
```bash
conda create -n llm python=3.9 # recommend to use 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 --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
```
### 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
```
```bash
python ./generate.py --prompt 'AI是什么？'
```
In the example, several arguments can be passed to satisfy your requirements:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM3 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
#### 2.3 Sample Output
#### [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b)
```log
Inference time: xxxx s
-------------------- Output --------------------
[gMASK]sop <|user|>
AI是什么？
<|assistant|> AI是人工智能(Artificial Intelligence)的缩写,指通过计算机程序或机器学习算法来模拟、延伸或扩展人类智能的技术。AI旨在
```
```log
Inference time: xxxx s
-------------------- Output --------------------
[gMASK]sop <|user|>
What is AI?
<|assistant|>
AI stands for Artificial Intelligence. It refers to the development of computer systems or machines that can perform tasks that would normally require human intelligence, such as recognizing patterns
```
## Example 2: Stream Chat using `stream_chat()` API
In the example [streamchat.py](./streamchat.py), we show a basic use case for a ChatGLM3 model to stream chat, 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](https://docs.conda.io/en/latest/miniconda.html#).
After installing conda, create a Python environment for BigDL-LLM:
```bash
conda create -n llm python=3.9 # recommend to use 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 --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
```
### 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
```
**Stream Chat using `stream_chat()` API**:
```
python ./streamchat.py
```
**Chat using `chat()` API**:
```
python ./streamchat.py --disable-stream
```
In the example, several arguments can be passed to satisfy your requirements:
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM3 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b'`.
- `--question QUESTION`: argument defining the question to ask. It is default to be `"晚上睡不着应该怎么办"`.
- `--disable-stream`: argument defining whether to stream chat. If include `--disable-stream` when running the script, the stream chat is disabled and `chat()` API is used.

74
python/llm/example/GPU/PyTorch-Models/Model/chatglm3/generate.py Normal file
View file

@ -0,0 +1,74 @@
#
# 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 intel_extension_for_pytorch as ipex
import time
import argparse
from transformers import AutoModel, AutoTokenizer
from bigdl.llm import optimize_model
# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM3 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--prompt', type=str, default="AI是什么？",
help='Prompt to infer')
parser.add_argument('--n-predict', type=int, default=32,
help='Max tokens to predict')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
# Load model
model = AutoModel.from_pretrained(model_path,
trust_remote_code=True,
torch_dtype='auto',
low_cpu_mem_usage=True)
# With only one line to enable BigDL-LLM optimization on model
model = optimize_model(model)
model = model.to('xpu')
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
# ipex model needs a warmup, then inference time can be accurate
output = model.generate(input_ids,
max_new_tokens=args.n_predict)
# start inference
st = time.time()
output = model.generate(input_ids,
max_new_tokens=args.n_predict)
torch.xpu.synchronize()
end = time.time()
output = output.cpu()
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
print(f'Inference time: {end-st} s')
print('-'*20, 'Output', '-'*20)
print(output_str)

75
python/llm/example/GPU/PyTorch-Models/Model/chatglm3/streamchat.py Normal file
View file

@ -0,0 +1,75 @@
#
# 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 intel_extension_for_pytorch as ipex
import time
import argparse
import numpy as np
from transformers import AutoModel, AutoTokenizer
from bigdl.llm import optimize_model
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Stream Chat for ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b",
help='The huggingface repo id for the ChatGLM3 model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--question', type=str, default="晚上睡不着应该怎么办",
help='Qustion you want to ask')
parser.add_argument('--disable-stream', action="store_true",
help='Disable stream chat')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
disable_stream = args.disable_stream
# Load model
model = AutoModel.from_pretrained(model_path,
trust_remote_code=True,
torch_dtype='auto',
low_cpu_mem_usage=True)
# With only one line to enable BigDL-LLM optimization on model
model = optimize_model(model)
model.to('xpu')
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
with torch.inference_mode():
prompt = args.question
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
# ipex model needs a warmup, then inference time can be accurate
output = model.generate(input_ids,
max_new_tokens=32)
# start inference
if disable_stream:
# Chat
response, history = model.chat(tokenizer, args.question, history=[])
print('-'*20, 'Chat Output', '-'*20)
print(response)
else:
# Stream chat
response_ = ""
print('-'*20, 'Stream Chat Output', '-'*20)
for response, history in model.stream_chat(tokenizer, args.question, history=[]):
print(response.replace(response_, ""), end="")
response_ = response