Skip to content

Commit c2f763d

Browse files
Oscilloscope98Oscilloscope98
authored
[LLM] Add more transformers int4 examples (Falcon) (intel#8546)
* Initial commit * Add Falcon examples and other small fix * Small fix * Small fix * Update based on comments * Small fix
1 parent 839fd70 commit c2f763d

File tree

5 files changed

+2492
-1
lines changed

5 files changed

+2492
-1
lines changed

‎python/llm/example/transformers/transformers_int4/baichuan/README.md

+1-1
Original file line numberDiff line numberDiff line change
@@ -23,7 +23,7 @@ python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROM
2323

2424
Arguments info:
2525
- `--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'`.
26-
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is AI?'`.
26+
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
2727
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
2828

2929
> **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.

‎python/llm/example/transformers/transformers_int4/falcon/README.md

+98
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,98 @@
1+
# Falcon
2+
3+
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.
4+
5+
## 0. Requirements
6+
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.
7+
8+
## Example: Predict Tokens using `generate()` API
9+
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.
10+
### 1. Install
11+
We suggest using conda to manage environment:
12+
```bash
13+
conda create -n llm python=3.9
14+
conda activate llm
15+
16+
pip install bigdl-llm[all] # install bigdl-llm with 'all' option
17+
pip install einops # additional package required for falcon-7b-instruct and falcon-40b-instruct to conduct generation
18+
```
19+
20+
### 2. (Optional) Download Model and Replace File
21+
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.
22+
23+
24+
#### 2.1 Download Model
25+
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.
26+
27+
```python
28+
from huggingface_hub import snapshot_download
29+
30+
# for tiiuae/falcon-7b-instruct
31+
model_path = snapshot_download(repo_id='tiiuae/falcon-7b-instruct',
32+
revision="c7f670a03d987254220f343c6b026ea0c5147185",
33+
cache_dir="dir/path/where/model/files/are/downloaded")
34+
print(f'tiiuae/falcon-7b-instruct checkpoint is downloaded to {model_path}')
35+
36+
# for tiiuae/falcon-40b-instruct
37+
model_path = snapshot_download(repo_id='tiiuae/falcon-40b-instruct',
38+
revision="1e7fdcc9f45d13704f3826e99937917e007cd975",
39+
cache_dir="dir/path/where/model/files/are/downloaded")
40+
print(f'tiiuae/falcon-40b-instruct checkpoint is downloaded to {model_path}')
41+
```
42+
43+
#### 2.2 Replace `modelling_RW.py`
44+
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).
45+
46+
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).
47+
48+
### 3. Run
49+
```
50+
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
51+
```
52+
53+
Arguments info:
54+
- `--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.
55+
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is AI?'`.
56+
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
57+
58+
> **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.
59+
>
60+
> Please select the appropriate size of the Falcon model based on the capabilities of your machine.
61+
62+
#### 3.1 Client
63+
On client Windows machine, it is recommended to run directly with full utilization of all cores:
64+
```powershell
65+
python ./generate.py
66+
```
67+
68+
#### 3.2 Server
69+
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.
70+
71+
E.g. on Linux,
72+
```bash
73+
# set BigDL-Nano env variables
74+
source bigdl-nano-init
75+
76+
# e.g. for a server with 48 cores per socket
77+
export OMP_NUM_THREADS=48
78+
numactl -C 0-47 -m 0 python ./generate.py
79+
```
80+
81+
#### 3.3 Sample Output
82+
#### [tiiuae/falcon-7b-instruct](https://huggingface.co/tiiuae/falcon-7b-instruct)
83+
```log
84+
Inference time: xxxx s
85+
-------------------- Prompt --------------------
86+
<human> What is AI? <bot>
87+
-------------------- Output --------------------
88+
<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?
89+
```
90+
91+
#### [tiiuae/falcon-40b-instruct](https://huggingface.co/tiiuae/falcon-40b-instruct)
92+
```log
93+
Inference time: xxxx s
94+
-------------------- Prompt --------------------
95+
<human> What is AI? <bot>
96+
-------------------- Output --------------------
97+
<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.
98+
```

0 commit comments

Comments
 (0)