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

LLM: add Modelscope model example (#10126)

Browse source
This commit is contained in:
binbin Deng 2024-02-08 11:18:07 +08:00 committed by GitHub
parent 0cf6a12691
commit 11fe5a87ec
6 changed files with 372 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

78
python/llm/example/CPU/ModelScope-Models/README.md Normal file
View file

@ -0,0 +1,78 @@
# Run ModelScope Model
In this directory, you will find example on how you could apply BigDL-LLM INT4 optimizations on ModelScope models. For illustration purposes, we utilize the [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary) as a reference ModelScope 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 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
pip install modelscope
```
### 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 ModelScope repo id for the ModelScope ChatGLM3 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be `'ZhipuAI/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-LLM env variables
source bigdl-llm-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
#### [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary)
```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
```

71
python/llm/example/CPU/ModelScope-Models/generate.py Normal file
View file

@ -0,0 +1,71 @@
#
# 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 modelscope 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 ModelScope ChatGLM3 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="ZhipuAI/chatglm3-6b",
help='The ModelScope repo id for the ChatGLM3 model to be downloaded'
', or the path to the ModelScope 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
# It is important to set `model_hub='modelscope'`, otherwise model hub is default to be huggingface
model = AutoModel.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True,
model_hub='modelscope')
# 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)

4
python/llm/example/CPU/README.md
View file

@ -5,11 +5,13 @@ This folder contains examples of running BigDL-LLM on Intel CPU:
- [HF-Transformers-AutoModels](HF-Transformers-AutoModels): running any ***Hugging Face Transformers*** model on BigDL-LLM (using the standard AutoModel APIs)
- [QLoRA-FineTuning](QLoRA-FineTuning): running ***QLoRA finetuning*** using BigDL-LLM on intel CPUs
- [vLLM-Serving](vLLM-Serving): running ***vLLM*** serving framework on intel CPUs (with BigDL-LLM low-bit optimized models)
- [Deepspeed-AutoTP](https://github.com/intel-analytics/BigDL/tree/main/python/llm/example/CPU/Deepspeed-AutoTP): running distributed inference using ***DeepSpeed AutoTP*** (with BigDL-LLM low-bit optimized models)
- [Deepspeed-AutoTP](Deepspeed-AutoTP): running distributed inference using ***DeepSpeed AutoTP*** (with BigDL-LLM low-bit optimized models)
- [LangChain](LangChain): running ***LangChain*** applications on BigDL-LLM
- [Applications](Applications): running LLM applications (such as agent, streaming-llm) on BigDl-LLM
- [PyTorch-Models](PyTorch-Models): running any PyTorch model on BigDL-LLM (with "one-line code change")
- [Native-Models](Native-Models): converting & running LLM in `llama`/`chatglm`/`bloom`/`gptneox`/`starcoder` model family using native (cpp) implementation
- [Speculative-Decoding](Speculative-Decoding): running any ***Hugging Face Transformers*** model with ***self-speculative decoding*** on Intel CPUs
- [ModelScope-Models](ModelScope-Models): running ***ModelScope*** model with BigDL-LLM on Intel CPUs
## System Support

135
python/llm/example/GPU/ModelScope-Models/README.md Normal file
View file

@ -0,0 +1,135 @@
# Run ModelScope Model
In this directory, you will find example on how you could apply BigDL-LLM INT4 optimizations on ModelScope models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary) as a reference ModelScope 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#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 on Intel GPUs.
### 1. Install
#### 1.1 Installation on Linux
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==2.1.10+xpu as default
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
pip install modelscope
```
#### 1.2 Installation on Windows
We suggest using conda to manage environment:
```bash
conda create -n llm python=3.9 libuv
conda activate llm
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
pip install modelscope
```
### 2. Configures OneAPI environment variables
#### 2.1 Configurations for Linux
```bash
source /opt/intel/oneapi/setvars.sh
```
#### 2.2 Configurations for Windows
```cmd
call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
```
> Note: Please make sure you are using **CMD** (**Anaconda Prompt** if using conda) to run the command as PowerShell is not supported.
### 3. Runtime Configurations
For optimal performance, it is recommended to set several environment variables. Please check out the suggestions based on your device.
#### 3.1 Configurations for Linux
<details>
<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
```bash
export USE_XETLA=OFF
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
```
</details>
<details>
<summary>For Intel Data Center GPU Max Series</summary>
```bash
export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export ENABLE_SDP_FUSION=1
```
> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
</details>
#### 3.2 Configurations for Windows
<details>
<summary>For Intel iGPU</summary>
```cmd
set SYCL_CACHE_PERSISTENT=1
set BIGDL_LLM_XMX_DISABLED=1
```
</details>
<details>
<summary>For Intel Arc™ A300-Series or Pro A60</summary>
```cmd
set SYCL_CACHE_PERSISTENT=1
```
</details>
<details>
<summary>For other Intel dGPU Series</summary>
There is no need to set further environment variables.
</details>
> Note: For the first time that each model runs on Intel iGPU/Intel Arc™ A300-Series or Pro A60, it may take several minutes to compile.
### 4. Running examples
```
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 ModelScope repo id for the ModelScope ChatGLM3 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be `'ZhipuAI/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
#### [ZhipuAI/chatglm3-6b](https://modelscope.cn/models/ZhipuAI/chatglm3-6b/summary)
```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
```

82
python/llm/example/GPU/ModelScope-Models/generate.py Normal file
View file

@ -0,0 +1,82 @@
#
# 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 modelscope 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="ZhipuAI/chatglm3-6b",
help='The ModelScope repo id for the ChatGLM3 model to be downloaded'
', or the path to the ModelScope 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
# When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
# This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
# It is important to set `model_hub='modelscope'`, otherwise model hub is default to be huggingface
model = AutoModel.from_pretrained(model_path,
load_in_4bit=True,
optimize_model=True,
trust_remote_code=True,
use_cache=True,
model_hub='modelscope')
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)

3
python/llm/example/GPU/README.md
View file

@ -2,11 +2,14 @@
This folder contains examples of running BigDL-LLM on Intel GPU:
- [Applications](Applications): running LLM applications (such as autogen) on BigDL-LLM
- [HF-Transformers-AutoModels](HF-Transformers-AutoModels): running any ***Hugging Face Transformers*** model on BigDL-LLM (using the standard AutoModel APIs)
- [LLM-Finetuning](LLM-Finetuning): running ***finetuning*** (such as LoRA, QLoRA, QA-LoRA, etc) using BigDL-LLM on Intel GPUs
- [vLLM-Serving](vLLM-Serving): running ***vLLM*** serving framework on intel GPUs (with BigDL-LLM low-bit optimized models)
- [Deepspeed-AutoTP](Deepspeed-AutoTP): running distributed inference using ***DeepSpeed AutoTP*** (with BigDL-LLM low-bit optimized models) on Intel GPUs
- [PyTorch-Models](PyTorch-Models): running any PyTorch model on BigDL-LLM (with "one-line code change")
- [Speculative-Decoding](Speculative-Decoding): running any ***Hugging Face Transformers*** model with ***self-speculative decoding*** on Intel GPUs
- [ModelScope-Models](ModelScope-Models): running ***ModelScope*** model with BigDL-LLM on Intel GPUs
## System Support