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

[LLM] Add more transformers int4 example (Llama 2) (#8602)

Browse source
This commit is contained in:
binbin Deng 2023-07-25 09:21:12 +08:00 committed by GitHub
parent 0f8201c730
commit 3f24202e4c
4 changed files with 159 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
python/llm/README.md
View file

@ -17,6 +17,7 @@ We may use any Hugging Face Transfomer models on `bigdl-llm`, and the following
| Model | Example |
|-----------|----------------------------------------------------------|
| LLaMA *(such as Vicuna, Guanaco, Koala, Baize, WizardLM, etc.)* | [link1](example/transformers/native_int4), [link2](example/transformers/transformers_int4/vicuna) |
| LLaMA 2 | [link](example/transformers/transformers_int4/llama2) |
| MPT | [link](example/transformers/transformers_int4/mpt) |
| Falcon | [link](example/transformers/transformers_int4/falcon) |
| ChatGLM | [link](example/transformers/transformers_int4/chatglm) |

1
python/llm/example/transformers/transformers_int4/README.md
View file

@ -5,6 +5,7 @@ You can use BigDL-LLM to run any Huggingface Transformer models with INT4 optimi
| Model | Example |
|-----------|----------------------------------------------------------|
| LLaMA | [link](vicuna) |
| LLaMA 2 | [link](llama2) |
| MPT | [link](mpt) |
| Falcon | [link](falcon) |
| ChatGLM | [link](chatglm) |

86
python/llm/example/transformers/transformers_int4/llama2/README.md Normal file
View file

@ -0,0 +1,86 @@
# Llama2
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Llama2 models. For illustration purposes, we utilize the [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, 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 Llama2 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 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'`.
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is AI?'`.
- `--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 Llama2 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
#### [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
### HUMAN:
What is AI?
### RESPONSE:
-------------------- Output --------------------
### HUMAN:
What is AI?
### RESPONSE:
AI is a term used to describe the development of computer systems that can perform tasks that typically require human intelligence, such as understanding natural language, recognizing images
```
#### [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
### HUMAN:
What is AI?
### RESPONSE:
-------------------- Output --------------------
### HUMAN:
What is AI?
### RESPONSE:
AI, or artificial intelligence, refers to the ability of machines to perform tasks that would typically require human intelligence, such as learning, problem-solving,
```

71
python/llm/example/transformers/transformers_int4/llama2/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
from bigdl.llm.transformers import AutoModelForCausalLM
from transformers import LlamaTokenizer
# 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
LLAMA2_PROMPT_FORMAT = """### HUMAN:
{prompt}
### RESPONSE:
"""
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
parser.add_argument('--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('--prompt', type=str, default="What is AI?",
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 = AutoModelForCausalLM.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True)
# Load tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = LLAMA2_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)