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

Add ziya CPU example (#10114)

Browse source 
* ziya on CPU

* add README for ziya

* specify use_cache

* add arc CPU

* update prompt format

* update link

* add comments to emphasize use_cache

* update pip cmd
This commit is contained in:
Zhicun 2024-02-20 13:59:52 +08:00 committed by GitHub
parent 71875ebc24
commit add3899311
6 changed files with 323 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
README.md
View file

@ -185,6 +185,7 @@ Over 40 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
| RWKV5 | | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/rwkv5) |
| Bark | [link](python/llm/example/CPU/PyTorch-Models/Model/bark) | [link](python/llm/example/GPU/PyTorch-Models/Model/bark) |
| SpeechT5 | | [link](python/llm/example/GPU/PyTorch-Models/Model/speech-t5) |
| Ziya-Coding-34B-v1.0 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya) | |
***For more details, please refer to the `bigdl-llm` [Document](https://test-bigdl-llm.readthedocs.io/en/main/doc/LLM/index.html), [Readme](python/llm), [Tutorial](https://github.com/intel-analytics/bigdl-llm-tutorial) and [API Doc](https://bigdl.readthedocs.io/en/latest/doc/PythonAPI/LLM/index.html).***

2
python/llm/README.md
View file

@ -81,7 +81,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
| RWKV5 | | [link](example/GPU/HF-Transformers-AutoModels/Model/rwkv5) |
| Bark | [link](example/CPU/PyTorch-Models/Model/bark) | [link](example/GPU/PyTorch-Models/Model/bark) |
| SpeechT5 | | [link](example/GPU/PyTorch-Models/Model/speech-t5) |
| Ziya-Coding-34B-v1.0 | [link](example/CPU/HF-Transformers-AutoModels/Model/ziya) | |
### Working with `bigdl-llm`

88
python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/README.md Normal file
View file

@ -0,0 +1,88 @@
# Ziya
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Ziya models. For illustration purposes, we utilize the [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) as a reference Ziya model.
> **Note**: If you want to download the Hugging Face *Transformers* model, please refer to [here](https://huggingface.co/docs/hub/models-downloading#using-git).
>
> BigDL-LLM optimizes the *Transformers* model in INT4 precision at runtime, and thus no explicit conversion is needed.
## 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 Ziya model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations.
### 1. Install
We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
After installing conda, create a Python environment for BigDL-LLM:
```bash
conda create -n llm python=3.9 # recommend to use Python 3.9
conda activate llm
pip install --pre --upgrade bigdl-llm[all] # install the latest bigdl-llm nightly build with 'all' option
pip install einops # additional package required for Ziya to conduct generation
```
### 2. Run
After setting up the Python environment, you could run the example by following steps.
> **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 Ziya model based on the capabilities of your machine.
#### 2.1 Client
On client Windows machines, it is recommended to run directly with full utilization of all cores:
```powershell
python ./generate.py --prompt 'def quick_sort(arr):\n'
```
More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
#### 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 --prompt 'def quick_sort(arr):\n'
```
More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
#### 2.3 Arguments Info
In the example, several arguments can be passed to satisfy your requirements:
- `--repo-id-or-model-path`: str, argument defining the huggingface repo id for the Ziya model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'IDEA-CCNL/Ziya-Coding-34B-v1.0'`.
- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `def quick_sort(arr):\n`.
- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `128`.
#### 2.4 Sample Output
#### [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
<human>:
def quick_sort(arr):\n
<bot>:
-------------------- Output --------------------
<s> <human>:
def quick_sort(arr):\n
<bot>:
def partition(arr, low, high):
i = (low-1)
pivot = arr[high]
for j in range(low, high):
if arr[j] <= pivot:
arr[i], arr[j] = arr[j], arr[i]
i = i+1
arr[i], arr[high] = arr[high], arr[i]
return i
def quick_sort(arr, low, high):
if low < high:
pi = partition(arr, low,
```

77
python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/generate.py Normal file
View file

@ -0,0 +1,77 @@
#
# 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 transformers import AutoTokenizer
# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0
ZIYA_PROMPT_FORMAT = "<human>: \n{prompt}\n<bot>: \n"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Ziya model')
parser.add_argument('--repo-id-or-model-path', type=str, default="IDEA-CCNL/Ziya-Coding-34B-v1.0",
help='The huggingface repo id for the Ziya model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--prompt', type=str, default="def quick_sort(arr):\n",
help='Prompt to infer')
parser.add_argument('--n-predict', type=int, default=128,
help='Max tokens to predict')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
from bigdl.llm.transformers import AutoModelForCausalLM
# 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 Ziya models
model = AutoModelForCausalLM.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True,
use_cache=True)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = ZIYA_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt")
st = time.time()
output = model.generate(input_ids,
max_new_tokens=args.n_predict,
do_sample = True,
top_p = 0.85,
temperature = 0.8,
repetition_penalty = 0.95,
eos_token_id = tokenizer.eos_token_id,
pad_token_id = tokenizer.pad_token_id,
)
end = time.time()
output_str = tokenizer.batch_decode(output)[0]
print(f'Inference time: {end-st} s')
print('-'*20, 'Prompt', '-'*20)
print(prompt)
print('-'*20, 'Output', '-'*20)
print(output_str)

78
python/llm/example/CPU/PyTorch-Models/Model/ziya/README.md Normal file
View file

@ -0,0 +1,78 @@
# Ziya
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate Ziya models. For illustration purposes, we utilize the [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) as a reference Ziya model.
## 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 Ziya model to predict the next N tokens using `generate()` API, with BigDL-LLM 'optimize_model' API.
### 1. Install
We suggest using conda to manage the Python environment. For more information about conda installation, please refer to [here](https://docs.conda.io/en/latest/miniconda.html#).
After installing conda, create a Python environment for BigDL-LLM:
```bash
conda create -n llm python=3.9 # recommend to use Python 3.9
conda activate llm
pip install --pre --upgrade bigdl-llm[all] # install the latest bigdl-llm nightly build with 'all' option
pip install einops # additional package required for Ziya to conduct generation
```
### 2. Run
After setting up the Python environment, you could run the example by following steps.
#### 2.1 Client
On client Windows machines, it is recommended to run directly with full utilization of all cores:
```powershell
python ./generate.py --prompt 'def quick_sort(arr):\n'
```
More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
#### 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
```
More information about arguments can be found in [Arguments Info](#23-arguments-info) section. The expected output can be found in [Sample Output](#24-sample-output) section.
#### 2.3 Arguments Info
In the example, several arguments can be passed to satisfy your requirements:
- `--repo-id-or-model-path`: str, argument defining the huggingface repo id for the Ziya model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'IDEA-CCNL/Ziya-Coding-34B-v1.0'`.
- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `def quick_sort(arr):\n`.
- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `128`.
#### 2.4 Sample Output
#### [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
<human>:
def quick_sort(arr):\n
<bot>:
-------------------- Output --------------------
<s> <human>:
def quick_sort(arr):\n
<bot>:
def partition(arr, low, high):
i = (low-1)
pivot = arr[high]
for j in range(low, high):
if arr[j] <= pivot:
arr[i], arr[j] = arr[j], arr[i]
i = i+1
arr[i], arr[high] = arr[high], arr[i]
return i
def quick_sort(arr, low, high):
if low < high:
pi = partition(arr, low,
```

78
python/llm/example/CPU/PyTorch-Models/Model/ziya/generate.py Normal file
View file

@ -0,0 +1,78 @@
#
# 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 transformers import AutoTokenizer
ZIYA_PROMPT_FORMAT = "<human>: \n{prompt}\n<bot>: \n"
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Ziya model')
parser.add_argument('--repo-id-or-model-path', type=str, default="IDEA-CCNL/Ziya-Coding-34B-v1.0",
help='The huggingface repo id for the Ziya model to be downloaded'
', or the path to the huggingface checkpoint folder')
parser.add_argument('--prompt', type=str, default="def quick_sort(arr):\n",
help='Prompt to infer')
parser.add_argument('--n-predict', type=int, default=128,
help='Max tokens to predict')
args = parser.parse_args()
model_path = args.repo_id_or_model_path
from transformers import AutoModelForCausalLM
from bigdl.llm import optimize_model
# 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 `optimization_model` API optimizations,
# it is important to set use_cache=True for Ziya models
model = AutoModelForCausalLM.from_pretrained(model_path,
trust_remote_code=True,
use_cache=True)
model = optimize_model(model)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path,
trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = ZIYA_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt")
st = time.time()
output = model.generate(input_ids,
max_new_tokens=args.n_predict,
do_sample = True,
top_p = 0.85,
temperature = 0.8,
repetition_penalty = 0.95,
eos_token_id = tokenizer.eos_token_id,
pad_token_id = tokenizer.pad_token_id,
)
end = time.time()
output_str = tokenizer.batch_decode(output)[0]
print(f'Inference time: {end-st} s')
print('-'*20, 'Prompt', '-'*20)
print(prompt)
print('-'*20, 'Output', '-'*20)
print(output_str)