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

[LLM] Add more transformers examples (ChatGLM) (#8521)

Browse source 
* Add example for chatglm v1 and other small fixes

* Small fix

* Small further fix

* Small fix

* Update based on comments & updates for client windows recommended settingts

* Small fix

* Small refactor

* Small fix

* Small fix

* Small fix to dolly v1

* Small fix
This commit is contained in:
Yuwen Hu 2023-07-14 16:41:13 +08:00 committed by GitHub
parent c87853233b
commit ca6e38607c
7 changed files with 178 additions and 47 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
python/llm/example/transformers/transformers_int4/README.md
View file

@ -6,12 +6,9 @@ To run the examples, we recommend using Intel® Xeon® processors (server), or >
For OS, BigDL-LLM supports Ubuntu 20.04 or later, CentOS 7 or later, and Windows 10/11.
## Best Known Configuration
For better performance, it is recommended to set environment variables with the help of BigDL-Nano:
## Best Known Configuration on Linux
For better performance, it is recommended to set environment variables on Linux with the help of BigDL-Nano:
```bash
pip install bigdl-nano
source bigdl-nano-init
```
following with
| Linux | Windows (powershell)|
|:------|:-------|
|`source bigdl-nano-init`|`bigdl-nano-init`|

72
python/llm/example/transformers/transformers_int4/chatglm/README.md Normal file
View file

@ -0,0 +1,72 @@
# ChatGLM
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on ChatGLM models. For illustration purposes, we utilize the [THUDM/chatglm-6b](https://huggingface.co/THUDM/chatglm-6b) as a reference ChatGLM 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 ChatGLM 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 --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 ChatGLM model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm-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 ChatGLM 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) 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/chatglm-6b](https://huggingface.co/THUDM/chatglm-6b)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
问：AI是什么？
答：
-------------------- Output --------------------
问:AI是什么?
答: AI是人工智能(Artificial Intelligence)的缩写,指的是一种能够模拟人类智能的技术或系统。AI系统可以通过学习、推理、解决问题等方式,实现类似于
```
```log
Inference time: xxxx s
-------------------- Prompt --------------------
问：What is AI?
答：
-------------------- Output --------------------
问:What is AI?
答: AI stands for "Artificial Intelligence." AI refers to the development of computer systems that can perform tasks that typically require human intelligence, such as recognizing speech, understanding natural
```

69
python/llm/example/transformers/transformers_int4/chatglm/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://huggingface.co/THUDM/chatglm-6b/blob/294cb13118a1e08ad8449ca542624a5c6aecc401/modeling_chatglm.py#L1281
CHATGLM_V1_PROMPT_FORMAT = "问：{prompt}\n答："
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM model')
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm-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 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_V1_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)

26
python/llm/example/transformers/transformers_int4/dolly/v1/README.md → python/llm/example/transformers/transformers_int4/dolly_v1/README.md
View file

@ -15,10 +15,7 @@ conda activate llm
pip install bigdl-llm[all] # install bigdl-llm with 'all' option
```
### 2. Config
It is recommended to set several environment variables for better performance. Please refer to [here](../README.md#best-known-configuration) for more information.
### 3. Run
### 2. Run
```
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
```
@ -32,27 +29,26 @@ Arguments info:
>
> Please select the appropriate size of the Dolly v1 model based on the capabilities of your machine.
#### 3.1 Client
For better utilization of multiple cores on the client machine, it is recommended to use all the performance-cores along with their hyperthreads.
E.g. on Windows,
#### 2.1 Client
On client Windows machine, it is recommended to run directly with full utilization of all cores:
```powershell
# for a client machine with 8 Performance-cores
$env:OMP_NUM_THREADS=16
python ./generate.py
```
#### 3.2 Server
On server, it is recommended to run the example with all the physical cores of a single socket.
#### 2.2 Server
For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration) for more information), and run the example with all the physical cores of a single socket.
E.g. on Linux,
```bash
# for a server with 48 cores per socket
# 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 -u ./generate.py
numactl -C 0-47 -m 0 python ./generate.py
```
#### 3.3 Sample Output
#### 2.3 Sample Output
#### [databricks/dolly-v1-6b](https://huggingface.co/databricks/dolly-v1-6b)
```log
Inference time: xxxx s

13
python/llm/example/transformers/transformers_int4/dolly/v1/generate.py → python/llm/example/transformers/transformers_int4/dolly_v1/generate.py
View file

@ -23,7 +23,7 @@ from bigdl.llm.transformers import AutoModelForCausalLM
from transformers import AutoTokenizer
# you could tune the prompt based on your own model,
# here the prompt format refers to https://huggingface.co/databricks/dolly-v1-6b#generate-text
# here the prompt tuning refers to https://huggingface.co/databricks/dolly-v1-6b#generate-text
DOLLY_V1_PROMPT_FORMAT = """Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
@ -33,7 +33,7 @@ DOLLY_V1_PROMPT_FORMAT = """Below is an instruction that describes a task. Write
"""
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Transformer INT4 example for Dolly v1 model')
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Dolly v1 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="databricks/dolly-v1-6b",
help='The huggingface repo id for the Dolly v1 model to be downloaded'
', or the path to the huggingface checkpoint folder')
@ -59,11 +59,12 @@ if __name__ == '__main__':
input_ids = tokenizer.encode(prompt, return_tensors="pt")
end_key_token_id=tokenizer.encode("### End")[0]
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
# enabling `use_cache=True` allows the model to utilize the previous
# key/values attentions to speed up decoding;
# to obtain optimal performance with BigDL-LLM INT4 optimizations,
# it is important to set use_cache=True for Dolly v1 models
output = model.generate(input_ids,
use_cache=True,
max_new_tokens=args.n_predict,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=end_key_token_id)

26
python/llm/example/transformers/transformers_int4/mpt/README.md
View file

@ -16,10 +16,7 @@ pip install bigdl-llm[all] # install bigdl-llm with 'all' option
pip install einops # additional package required for mpt-7b-chat to conduct generation
```
### 2. Config
It is recommended to set several environment variables for better performance. Please refer to [here](../README.md#best-known-configuration) for more information.
### 3. Run
### 2. Run
```
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
```
@ -33,27 +30,26 @@ Arguments info:
>
> Please select the appropriate size of the MPT model based on the capabilities of your machine.
#### 3.1 Client
For better utilization of multiple cores on the client machine, it is recommended to use all the performance-cores along with their hyperthreads.
E.g. on Windows,
#### 2.1 Client
On client Windows machine, it is recommended to run directly with full utilization of all cores:
```powershell
# for a client machine with 8 Performance-cores
$env:OMP_NUM_THREADS=16
python ./generate.py
```
#### 3.2 Server
On server, it is recommended to run the example with all the physical cores of a single socket.
#### 2.2 Server
For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration) for more information), and run the example with all the physical cores of a single socket.
E.g. on Linux,
```bash
# for a server with 48 cores per socket
# 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 -u ./generate.py
numactl -C 0-47 -m 0 python ./generate.py
```
#### 3.3 Sample Output
#### 2.3 Sample Output
#### [mosaicml/mpt-7b-chat](https://huggingface.co/mosaicml/mpt-7b-chat)
```log
Inference time: xxxx s

6
python/llm/example/transformers/transformers_int4/mpt/generate.py
View file

@ -25,7 +25,7 @@ from transformers import AutoTokenizer
MPT_PROMPT_FORMAT = "<human>{prompt} <bot>"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Transformer INT4 example for MPT model')
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for MPT model')
parser.add_argument('--repo-id-or-model-path', type=str, default="mosaicml/mpt-7b-chat",
help='The huggingface repo id for the MPT to be downloaded'
', or the path to the huggingface checkpoint folder')
@ -40,8 +40,8 @@ if __name__ == '__main__':
# Load model in 4 bit,
# which convert the relevant layers in the model into INT4 format
model = AutoModelForCausalLM.from_pretrained(model_path,
trust_remote_code=True,
load_in_4bit=True)
load_in_4bit=True,
trust_remote_code=True)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,