LLM: add CodeLlama CPU and GPU examples (#9338)
* LLM: add codellama CPU pytorch examples * LLM: add codellama CPU transformers examples * LLM: add codellama GPU transformers examples * LLM: add codellama GPU pytorch examples * LLM: add codellama in readme * LLM: add LLaVA link
This commit is contained in:
Jin Qiao
GitHub
parent
d4dffbdb62
commit
97a38958bd
10 changed files with 531 additions and 2 deletions
|
|
@ -156,10 +156,12 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
|
|||
| Phi-1_5 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5) |
|
||||
| Flan-t5 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5) |
|
||||
| Qwen-VL | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/qwen-vl) | |
|
||||
| LLaVA | [link](python/llm/example/CPU/PyTorch-Models/Model/llava) | [link](python/llm/example/GPU/PyTorch-Models/Model/llava) |
|
||||
| LLaVA | [link](python/llm/example/CPU/PyTorch-Models/Model/llava) | [link](python/llm/example/GPU/PyTorch-Models/Model/llava) |
|
||||
| CodeLlama | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/codellama) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/codellama) |
|
||||
| Skywork | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/skywork) | |
|
||||
|
||||
|
||||
|
||||
***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).***
|
||||
|
||||
---
|
||||
|
|
|
|||
|
|
@ -64,7 +64,9 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa
|
|||
| Flan-t5 | [link](example/CPU/HF-Transformers-AutoModels/Model/flan-t5) | [link](example/GPU/HF-Transformers-AutoModels/Model/flan-t5) |
|
||||
| Qwen-VL | [link](example/CPU/HF-Transformers-AutoModels/Model/qwen-vl) | |
|
||||
| LLaVA | [link](example/CPU/PyTorch-Models/Model/llava) | [link](example/GPU/PyTorch-Models/Model/llava) |
|
||||
| Skywork | [link](example/CPU/HF-Transformers-AutoModels/Model/skywork) | |
|
||||
| CodeLlama | [link](example/CPU/HF-Transformers-AutoModels/Model/codellama) | [link](example/GPU/HF-Transformers-AutoModels/Model/codellama) |
|
||||
| Skywork | [link](example/CPU/HF-Transformers-AutoModels/Model/skywork) | |
|
||||
|
||||
|
||||
### Working with `bigdl-llm`
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,66 @@
|
|||
# CodeLlama
|
||||
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on CodeLlama models. For illustration purposes, we utilize the [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf) as reference CodeLlama 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 CodeLlama 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 CodeLlama model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'codellama/CodeLlama-7b-hf'`.
|
||||
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'def print_hello_world():'`.
|
||||
- `--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 CodeLlama 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
|
||||
#### [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
def print_hello_world():
|
||||
<FILL_ME>
|
||||
-------------------- Output --------------------
|
||||
def print_hello_world():
|
||||
print("Hello World")
|
||||
|
||||
|
||||
def print_hello_world_with_name(name):
|
||||
print("Hello " + name)
|
||||
|
||||
```
|
||||
|
|
@ -0,0 +1,67 @@
|
|||
#
|
||||
# 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 CodeLlamaTokenizer
|
||||
|
||||
# you could tune the prompt based on your own model,
|
||||
# here the prompt tuning refers to https://huggingface.co/docs/transformers/v4.34.1/model_doc/code_llama
|
||||
CODELLAMA_PROMPT_FORMAT = "{prompt}\n<FILL_ME>"
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeLlama model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="codellama/CodeLlama-7b-hf",
|
||||
help='The huggingface repo id for the CodeLlama to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--prompt', type=str, default="def print_hello_world():",
|
||||
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 = CodeLlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
prompt = CODELLAMA_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)
|
||||
|
|
@ -0,0 +1,63 @@
|
|||
# CodeLlama
|
||||
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate CodeLlama models. For illustration purposes, we utilize the [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf) as reference CodeLlama models.
|
||||
|
||||
## 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 CodeLlama 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
|
||||
```
|
||||
|
||||
### 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 print_hello_world():'
|
||||
```
|
||||
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-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 --prompt 'def print_hello_world():'
|
||||
```
|
||||
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 REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the CodeLlama model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'codellama/CodeLlama-7b-hf'`.
|
||||
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'def print_hello_world():'`.
|
||||
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
#### 2.3 Sample Output
|
||||
#### [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Output --------------------
|
||||
def print_hello_world():
|
||||
print("Hello World")
|
||||
|
||||
|
||||
def print_hello_world_with_name(name):
|
||||
print("Hello " + name)
|
||||
|
||||
```
|
||||
|
|
@ -0,0 +1,61 @@
|
|||
#
|
||||
# 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 transformers import AutoModelForCausalLM, CodeLlamaTokenizer
|
||||
from bigdl.llm import optimize_model
|
||||
|
||||
# you could tune the prompt based on your own model,
|
||||
# here the prompt tuning refers to https://huggingface.co/docs/transformers/v4.34.1/model_doc/code_llama
|
||||
CODELLAMA_PROMPT_FORMAT = "{prompt}\n<FILL_ME>"
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeLlama model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="codellama/CodeLlama-7b-hf",
|
||||
help='The huggingface repo id for the CodeLlama to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--prompt', type=str, default="def print_hello_world():",
|
||||
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
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# With only one line to enable BigDL-LLM optimization on model
|
||||
model = optimize_model(model)
|
||||
|
||||
# Load tokenizer
|
||||
tokenizer = CodeLlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
prompt = CODELLAMA_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)
|
||||
end = time.time()
|
||||
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
|
||||
print(f'Inference time: {end-st} s')
|
||||
print('-'*20, 'Output', '-'*20)
|
||||
print(output_str)
|
||||
|
|
@ -0,0 +1,79 @@
|
|||
#
|
||||
# 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 intel_extension_for_pytorch as ipex
|
||||
import time
|
||||
import argparse
|
||||
|
||||
from bigdl.llm.transformers import AutoModelForCausalLM
|
||||
from transformers import CodeLlamaTokenizer
|
||||
|
||||
# you could tune the prompt based on your own model,
|
||||
# here the prompt tuning refers to https://huggingface.co/docs/transformers/v4.34.1/model_doc/code_llama
|
||||
CODELLAMA_PROMPT_FORMAT = "{prompt}\n<FILL_ME>"
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeLlama model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="codellama/CodeLlama-7b-hf",
|
||||
help='The huggingface repo id for the CodeLlama to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--prompt', type=str, default="def print_hello_world():",
|
||||
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,
|
||||
optimize_model=False,
|
||||
trust_remote_code=True,
|
||||
use_cache=True)
|
||||
model = model.to('xpu')
|
||||
|
||||
# Load tokenizer
|
||||
tokenizer = CodeLlamaTokenizer.from_pretrained(model_path,
|
||||
trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
prompt = CODELLAMA_PROMPT_FORMAT.format(prompt=args.prompt)
|
||||
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
|
||||
|
||||
# ipex model needs a warmup, then inference time can be accurate
|
||||
output = model.generate(input_ids,
|
||||
max_new_tokens=args.n_predict)
|
||||
|
||||
# start inference
|
||||
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)
|
||||
torch.xpu.synchronize()
|
||||
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)
|
||||
|
|
@ -0,0 +1,56 @@
|
|||
# CodeLlama
|
||||
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on CodeLlama models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf) as a reference CodeLlama model.
|
||||
|
||||
## 0. Requirements
|
||||
To run these examples with BigDL-LLM on Intel GPUs, 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 an CodeLlama model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs.
|
||||
### 1. Install
|
||||
We suggest using conda to manage environment:
|
||||
```bash
|
||||
conda create -n llm python=3.9
|
||||
conda activate llm
|
||||
# below command will install intel_extension_for_pytorch==2.0.110+xpu as default
|
||||
# you can install specific ipex/torch version for your need
|
||||
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
|
||||
```
|
||||
|
||||
### 2. Configures OneAPI environment variables
|
||||
```bash
|
||||
source /opt/intel/oneapi/setvars.sh
|
||||
```
|
||||
|
||||
### 3. Run
|
||||
|
||||
For optimal performance on Arc, it is recommended to set several environment variables.
|
||||
|
||||
```bash
|
||||
export USE_XETLA=OFF
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
```
|
||||
|
||||
```
|
||||
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 CodeLlama model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'codellama/CodeLlama-7b-hf'`.
|
||||
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'def print_hello_world():'`.
|
||||
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
#### Sample Output
|
||||
#### [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
def print_hello_world():
|
||||
<FILL_ME>
|
||||
-------------------- Output --------------------
|
||||
def print_hello_world():
|
||||
print("Hello World!")
|
||||
|
||||
|
||||
def print_hello_world_with_args(name):
|
||||
print(f"Hello {name
|
||||
```
|
||||
|
|
@ -0,0 +1,58 @@
|
|||
# CodeLlama
|
||||
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate CodeLlama models. For illustration purposes, we utilize the [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf) as reference CodeLlama models.
|
||||
|
||||
## Requirements
|
||||
To run these examples with BigDL-LLM on Intel GPUs, 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 CodeLlama model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs.
|
||||
### 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
|
||||
|
||||
# below command will install intel_extension_for_pytorch==2.0.110+xpu as default
|
||||
# you can install specific ipex/torch version for your need
|
||||
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
|
||||
```
|
||||
|
||||
### 2. Configures OneAPI environment variables
|
||||
```bash
|
||||
source /opt/intel/oneapi/setvars.sh
|
||||
```
|
||||
|
||||
### 3. Run
|
||||
|
||||
For optimal performance on Arc, it is recommended to set several environment variables.
|
||||
|
||||
```bash
|
||||
export USE_XETLA=OFF
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
```
|
||||
|
||||
```bash
|
||||
python ./generate.py --prompt 'def print_hello_world():'
|
||||
```
|
||||
|
||||
In the example, several arguments can be passed to satisfy your requirements:
|
||||
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the CodeLlama model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'codellama/CodeLlama-7b-hf'`.
|
||||
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `def print_hello_world():'`.
|
||||
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
#### 2.3 Sample Output
|
||||
#### [codellama/CodeLlama-7b-hf](https://huggingface.co/codellama/CodeLlama-7b-hf)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Output --------------------
|
||||
def print_hello_world():
|
||||
print("Hello World!")
|
||||
print_hello_world()
|
||||
def print_hello_world():
|
||||
print("Hello World!")
|
||||
|
||||
|
||||
```
|
||||
|
|
@ -0,0 +1,75 @@
|
|||
#
|
||||
# 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 intel_extension_for_pytorch as ipex
|
||||
import time
|
||||
import argparse
|
||||
|
||||
from transformers import AutoModelForCausalLM, CodeLlamaTokenizer
|
||||
from bigdl.llm import optimize_model
|
||||
|
||||
# you could tune the prompt based on your own model,
|
||||
# here the prompt tuning refers to https://huggingface.co/docs/transformers/v4.34.1/model_doc/code_llama
|
||||
CODELLAMA_PROMPT_FORMAT = "{prompt}\n<FILL_ME>"
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for CodeLlama model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="codellama/CodeLlama-7b-hf",
|
||||
help='The huggingface repo id for the CodeLlama model to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--prompt', type=str, default="def print_hello_world():",
|
||||
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
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path,
|
||||
trust_remote_code=True,
|
||||
torch_dtype='auto',
|
||||
low_cpu_mem_usage=True)
|
||||
|
||||
# With only one line to enable BigDL-LLM optimization on model
|
||||
model = optimize_model(model)
|
||||
|
||||
model = model.to('xpu')
|
||||
|
||||
# Load tokenizer
|
||||
tokenizer = CodeLlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
prompt = CODELLAMA_PROMPT_FORMAT.format(prompt=args.prompt)
|
||||
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
|
||||
# ipex model needs a warmup, then inference time can be accurate
|
||||
output = model.generate(input_ids,
|
||||
max_new_tokens=args.n_predict)
|
||||
|
||||
# start inference
|
||||
st = time.time()
|
||||
output = model.generate(input_ids,
|
||||
max_new_tokens=args.n_predict)
|
||||
torch.xpu.synchronize()
|
||||
end = time.time()
|
||||
output = output.cpu()
|
||||
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
|
||||
print(f'Inference time: {end-st} s')
|
||||
print('-'*20, 'Output', '-'*20)
|
||||
print(output_str)
|
||||
Loading…
Reference in a new issue