[LLM] Add more transformers int4 examples (Falcon) (#8546)
* Initial commit * Add Falcon examples and other small fix * Small fix * Small fix * Update based on comments * Small fix
This commit is contained in:
Yuwen Hu
GitHub
parent
de772e7a80
commit
1344f50f75
5 changed files with 2492 additions and 1 deletions
|
|
@ -23,7 +23,7 @@ python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROM
|
|||
|
||||
Arguments info:
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Baichuan model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'baichuan-inc/Baichuan-13B-Chat'`.
|
||||
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is AI?'`.
|
||||
- `--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.
|
||||
|
|
|
|||
|
|
@ -0,0 +1,98 @@
|
|||
# Falcon
|
||||
|
||||
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Falcon models. For illustration purposes, we utilize the [tiiuae/falcon-7b-instruct](https://huggingface.co/tiiuae/falcon-7b-instruct) and [tiiuae/falcon-40b-instruct](https://huggingface.co/tiiuae/falcon-40b-instruct) as reference Falcon 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 Falcon 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
|
||||
pip install einops # additional package required for falcon-7b-instruct and falcon-40b-instruct to conduct generation
|
||||
```
|
||||
|
||||
### 2. (Optional) Download Model and Replace File
|
||||
If you select the Falcon models ([tiiuae/falcon-7b-instruct](https://huggingface.co/tiiuae/falcon-7b-instruct) or [tiiuae/falcon-40b-instruct](https://huggingface.co/tiiuae/falcon-40b-instruct)), please note that their code (`modelling_RW.py`) does not support KV cache at the moment. To address issue, we have provided two updated files ([falcon-7b-instruct/modelling_RW.py](./falcon-7b-instruct/modelling_RW.py) and [falcon-40b-instruct/modelling_RW.py](./falcon-40b-instruct/modelling_RW.py)), which can be used to achieve the best performance using BigDL-LLM INT4 optimizations with KV cache support.
|
||||
|
||||
|
||||
#### 2.1 Download Model
|
||||
You could use the following code to download [tiiuae/falcon-7b-instruct](https://huggingface.co/tiiuae/falcon-7b-instruct) or [tiiuae/falcon-40b-instruct](https://huggingface.co/tiiuae/falcon-40b-instruct) with a specific snapshot id. Please note that the `modelling_RW.py` files that we provide are based on these specific commits.
|
||||
|
||||
```python
|
||||
from huggingface_hub import snapshot_download
|
||||
|
||||
# for tiiuae/falcon-7b-instruct
|
||||
model_path = snapshot_download(repo_id='tiiuae/falcon-7b-instruct',
|
||||
revision="c7f670a03d987254220f343c6b026ea0c5147185",
|
||||
cache_dir="dir/path/where/model/files/are/downloaded")
|
||||
print(f'tiiuae/falcon-7b-instruct checkpoint is downloaded to {model_path}')
|
||||
|
||||
# for tiiuae/falcon-40b-instruct
|
||||
model_path = snapshot_download(repo_id='tiiuae/falcon-40b-instruct',
|
||||
revision="1e7fdcc9f45d13704f3826e99937917e007cd975",
|
||||
cache_dir="dir/path/where/model/files/are/downloaded")
|
||||
print(f'tiiuae/falcon-40b-instruct checkpoint is downloaded to {model_path}')
|
||||
```
|
||||
|
||||
#### 2.2 Replace `modelling_RW.py`
|
||||
For `tiiuae/falcon-7b-instruct`, you should replace the `modelling_RW.py` with [falcon-7b-instruct/modelling_RW.py](./falcon-7b-instruct/modelling_RW.py).
|
||||
|
||||
For `tiiuae/falcon-40b-instruct`, you should replace the `modelling_RW.py` with [falcon-40b-instruct/modelling_RW.py](./falcon-40b-instruct/modelling_RW.py).
|
||||
|
||||
### 3. 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 Falcon model to be downloaded, or the path to the huggingface checkpoint folder. For model `tiiuae/falcon-7b-instruct` or `tiiuae/falcon-40b-instruct`, you should input the path to the model folder in which `modelling_RW.py` has been replaced.
|
||||
- `--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 Falcon model based on the capabilities of your machine.
|
||||
|
||||
#### 3.1 Client
|
||||
On client Windows machine, it is recommended to run directly with full utilization of all cores:
|
||||
```powershell
|
||||
python ./generate.py
|
||||
```
|
||||
|
||||
#### 3.2 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
|
||||
```
|
||||
|
||||
#### 3.3 Sample Output
|
||||
#### [tiiuae/falcon-7b-instruct](https://huggingface.co/tiiuae/falcon-7b-instruct)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
<human> What is AI? <bot>
|
||||
-------------------- Output --------------------
|
||||
<human> What is AI? <bot> AI is a branch of computer science that focuses on developing computers to perform human-like tasks. <human> What are some examples of these tasks?
|
||||
```
|
||||
|
||||
#### [tiiuae/falcon-40b-instruct](https://huggingface.co/tiiuae/falcon-40b-instruct)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
<human> What is AI? <bot>
|
||||
-------------------- Output --------------------
|
||||
<human> What is AI? <bot> AI stands for Artificial Intelligence. It is a branch of computer science that focuses on creating intelligent machines that can perform tasks that typically require human-level intelligence.
|
||||
```
|
||||
File diff suppressed because it is too large
Load diff
File diff suppressed because it is too large
Load diff
|
|
@ -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
|
||||
|
||||
from bigdl.llm.transformers import AutoModelForCausalLM
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
# you could tune the prompt based on your own model,
|
||||
FALCON_PROMPT_FORMAT = "<human> {prompt} <bot>"
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Falcon model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str,
|
||||
help='The huggingface repo id for the Falcon model to be downloaded, '
|
||||
'or the path to the huggingface checkpoint folder. '
|
||||
'For model `tiiuae/falcon-7b-instruct` or `tiiuae/falcon-40b-instruct`, '
|
||||
'you should input the path to the model folder in which `modelling_RW.py` has been replaced')
|
||||
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 = AutoTokenizer.from_pretrained(model_path,
|
||||
trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
prompt = FALCON_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)
|
||||
Loading…
Reference in a new issue