Add chatglm3 long input example (#10739)
* Add long context input example for chatglm3 * Small fix * Small fix * Small fix
This commit is contained in:
Keyan (Kyrie) Zhang
GitHub
parent
fd473ddb1b
commit
1256a2cc4e
5 changed files with 224 additions and 8 deletions
1
python/llm/example/GPU/Long-Context/Chatglm3-32K/8k.txt
Normal file
1
python/llm/example/GPU/Long-Context/Chatglm3-32K/8k.txt
Normal file
File diff suppressed because one or more lines are too long
126
python/llm/example/GPU/Long-Context/Chatglm3-32K/README.md
Normal file
126
python/llm/example/GPU/Long-Context/Chatglm3-32K/README.md
Normal file
|
|
@ -0,0 +1,126 @@
|
|||
# Chatglm3-32k
|
||||
In this directory, you will find examples on how you could apply IPEX-LLM INT4/FP8 optimizations on Chatglm3-32K models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [THUDM/chatglm3-6b-32k](https://huggingface.co/THUDM/chatglm3-6b-32k) as reference Chatglm3-32K models.
|
||||
|
||||
## 0. Requirements
|
||||
To run these examples with IPEX-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../../../README.md#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 Chatglm3 model to predict the next N tokens using `generate()` API, with IPEX-LLM INT4/FP8 optimizations on Intel GPUs.
|
||||
### 1. Install
|
||||
#### 1.1 Installation on Linux
|
||||
We suggest using conda to manage environment:
|
||||
```bash
|
||||
conda create -n llm python=3.11
|
||||
conda activate llm
|
||||
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
|
||||
pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
|
||||
```
|
||||
|
||||
#### 1.2 Installation on Windows
|
||||
We suggest using conda to manage environment:
|
||||
```bash
|
||||
conda create -n llm python=3.11 libuv
|
||||
conda activate llm
|
||||
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
|
||||
pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
|
||||
```
|
||||
|
||||
### 2. Configures OneAPI environment variables
|
||||
#### 2.1 Configurations for Linux
|
||||
```bash
|
||||
source /opt/intel/oneapi/setvars.sh
|
||||
```
|
||||
#### 2.2 Configurations for Windows
|
||||
```cmd
|
||||
call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
|
||||
```
|
||||
> Note: Please make sure you are using **CMD** (**Anaconda Prompt** if using conda) to run the command as PowerShell is not supported.
|
||||
### 3. Runtime Configurations
|
||||
For optimal performance, it is recommended to set several environment variables. Please check out the suggestions based on your device.
|
||||
#### 3.1 Configurations for Linux
|
||||
<details>
|
||||
|
||||
<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
|
||||
|
||||
```bash
|
||||
export USE_XETLA=OFF
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
<details>
|
||||
|
||||
<summary>For Intel Data Center GPU Max Series</summary>
|
||||
|
||||
```bash
|
||||
export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
export ENABLE_SDP_FUSION=1
|
||||
```
|
||||
> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
|
||||
</details>
|
||||
|
||||
#### 3.2 Configurations for Windows
|
||||
<details>
|
||||
|
||||
<summary>For Intel iGPU</summary>
|
||||
|
||||
```cmd
|
||||
set SYCL_CACHE_PERSISTENT=1
|
||||
set BIGDL_LLM_XMX_DISABLED=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
<details>
|
||||
|
||||
<summary>For Intel Arc™ A300-Series or Pro A60</summary>
|
||||
|
||||
```cmd
|
||||
set SYCL_CACHE_PERSISTENT=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
<details>
|
||||
|
||||
<summary>For other Intel dGPU Series</summary>
|
||||
|
||||
There is no need to set further environment variables.
|
||||
|
||||
</details>
|
||||
|
||||
> Note: For the first time that each model runs on Intel iGPU/Intel Arc™ A300-Series or Pro A60, it may take several minutes to compile.
|
||||
### 4. Running examples
|
||||
#### 4.1 Using simple prompt
|
||||
```
|
||||
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT --low-bit LOW_BIT
|
||||
```
|
||||
|
||||
Arguments info:
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Chatglm3 model (e.g. `THUDM/chatglm3-6b-32k`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b-32k'`.
|
||||
- `--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`.
|
||||
- `--low-bit LOW_BIT`: argument defining which low bit optimization to use. Options are sym_int4 or fp8. It is default to be `sym_int4`.
|
||||
|
||||
#### 4.2 Using long context input prompt
|
||||
You can set the `prompt` argument to be a `.txt` file path containing the long context prompt text. An example command using the 8k input size prompt we provide is given below:
|
||||
```
|
||||
python ./generate.py --repo-id-or-model-path togethercomputer/chatglm3-6b-32k --prompt 8k.txt
|
||||
```
|
||||
> Note: If you need to run longer input or use less memory, please set `IPEX_LLM_LOW_MEM=1`, which will enable memory optimization and may slightly affect the latency performance.
|
||||
#### Sample Output
|
||||
#### [THUDM/chatglm3-6b-32k](https://huggingface.co/THUDM/chatglm3-6b-32k)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
<|user|>
|
||||
What is AI?
|
||||
<|assistant|>
|
||||
-------------------- Output --------------------
|
||||
[gMASK]sop <|user|>
|
||||
What is AI?
|
||||
<|assistant|>
|
||||
AI stands for Artificial Intelligence. It refers to the ability of computers and other machines to perform tasks that typically require human intelligence, such as recognizing patterns, making
|
||||
```
|
||||
86
python/llm/example/GPU/Long-Context/Chatglm3-32K/generate.py
Normal file
86
python/llm/example/GPU/Long-Context/Chatglm3-32K/generate.py
Normal file
|
|
@ -0,0 +1,86 @@
|
|||
#
|
||||
# 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 ipex_llm.transformers import AutoModel
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
# you could tune the prompt based on your own model,
|
||||
# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
|
||||
CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b-32k",
|
||||
help='The huggingface repo id for the ChatGLM3 model 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')
|
||||
parser.add_argument('--low-bit', type=str, default="sym_int4",
|
||||
help='Load model in low bit')
|
||||
|
||||
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
|
||||
# When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
|
||||
# This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
|
||||
model = AutoModel.from_pretrained(model_path,
|
||||
load_in_low_bit=args.low_bit,
|
||||
optimize_model=True,
|
||||
trust_remote_code=True,
|
||||
use_cache=True)
|
||||
model = model.half().to('xpu')
|
||||
|
||||
# Load tokenizer
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path,
|
||||
trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
if not args.prompt.endswith('.txt'):
|
||||
prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
|
||||
else:
|
||||
with open(args.prompt, 'r') as f:
|
||||
prompt = f.read()
|
||||
input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
|
||||
# ipex_llm 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 IPEX-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)
|
||||
|
|
@ -1,11 +1,11 @@
|
|||
# Llama2
|
||||
In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on Llama2-32K models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [togethercomputer/Llama-2-7B-32K-Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct) as reference Llama2-32K models.
|
||||
# Llama2-32k
|
||||
In this directory, you will find examples on how you could apply IPEX-LLM INT4/FP8 optimizations on Llama2-32K models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [togethercomputer/Llama-2-7B-32K-Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct) as reference Llama2-32K models.
|
||||
|
||||
## 0. Requirements
|
||||
To run these examples with IPEX-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../../../README.md#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 IPEX-LLM INT4 optimizations on Intel GPUs.
|
||||
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 IPEX-LLM INT4/FP8 optimizations on Intel GPUs.
|
||||
### 1. Install
|
||||
#### 1.1 Installation on Linux
|
||||
We suggest using conda to manage environment:
|
||||
|
|
@ -95,20 +95,21 @@ There is no need to set further environment variables.
|
|||
### 4. Running examples
|
||||
#### 4.1 Using simple prompt
|
||||
```
|
||||
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
|
||||
python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT --low-bit LOW_BIT
|
||||
```
|
||||
|
||||
Arguments info:
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Llama2 model (e.g. `togethercomputer/Llama-2-7B-32K-Instruct`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'togethercomputer/Llama-2-7B-32K-Instruct'`.
|
||||
- `--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`.
|
||||
- `--low-bit LOW_BIT`: argument defining which low bit optimization to use. Options are sym_int4 or fp8. It is default to be `sym_int4`.
|
||||
|
||||
#### 4.2 Using 8k input size prompt
|
||||
You can set the `prompt` argument to be a `.txt` file path containing the 8k size prompt text. An example command using the 8k input size prompt we provide is given below:
|
||||
#### 4.2 Using long context input prompt
|
||||
You can set the `prompt` argument to be a `.txt` file path containing the long context prompt text. An example command using the 8k input size prompt we provide is given below:
|
||||
```
|
||||
python ./generate.py --repo-id-or-model-path togethercomputer/Llama-2-7B-32K-Instruct --prompt 8k.txt
|
||||
```
|
||||
> Note: If you need to use less memory, please set `IPEX_LLM_LOW_MEM=1`, which will enable memory optimization and may slightly affect the latency performance.
|
||||
> Note: If you need to run longer input or use less memory, please set `IPEX_LLM_LOW_MEM=1`, which will enable memory optimization and may slightly affect the latency performance.
|
||||
#### Sample Output
|
||||
#### [togethercomputer/Llama-2-7B-32K-Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct)
|
||||
```log
|
||||
|
|
|
|||
|
|
@ -48,6 +48,8 @@ if __name__ == '__main__':
|
|||
help='Prompt to infer')
|
||||
parser.add_argument('--n-predict', type=int, default=32,
|
||||
help='Max tokens to predict')
|
||||
parser.add_argument('--low-bit', type=str, default="sym_int4",
|
||||
help='Load model in low bit')
|
||||
|
||||
args = parser.parse_args()
|
||||
model_path = args.repo_id_or_model_path
|
||||
|
|
@ -57,7 +59,7 @@ if __name__ == '__main__':
|
|||
# When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
|
||||
# This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path,
|
||||
load_in_4bit=True,
|
||||
load_in_low_bit=args.low_bit,
|
||||
optimize_model=True,
|
||||
trust_remote_code=True,
|
||||
use_cache=True)
|
||||
|
|
|
|||
Loading…
Reference in a new issue