Add DeepSeek-MoE-16B-Chat (#10155)
* dsmoe-hf add * add dsmoe pytorch * update README * modify comment * remove GPU example * update model name * format code
This commit is contained in:
Zhicun
GitHub
parent
f4a2e32106
commit
308e637d0d
7 changed files with 264 additions and 0 deletions
|
|
@ -189,12 +189,14 @@ 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) |
|
||||
| DeepSeek-MoE | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe) | |
|
||||
| Ziya-Coding-34B-v1.0 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya) | |
|
||||
| Phi-2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2) |
|
||||
| Yuan2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/yuan2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/yuan2) |
|
||||
| Gemma | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/gemma) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/gemma) |
|
||||
| DeciLM-7B | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/deciLM-7b) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/deciLM-7b) |
|
||||
|
||||
|
||||
***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).***
|
||||
|
||||
---
|
||||
|
|
|
|||
|
|
@ -81,6 +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) |
|
||||
| DeepSeek-MoE | [link](example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe) | |
|
||||
| Ziya-Coding-34B-v1.0 | [link](example/CPU/HF-Transformers-AutoModels/Model/ziya) | |
|
||||
| Phi-2 | [link](example/CPU/HF-Transformers-AutoModels/Model/phi-2) | [link](example/GPU/HF-Transformers-AutoModels/Model/phi-2) |
|
||||
| Yuan2 | [link](example/CPU/HF-Transformers-AutoModels/Model/yuan2) | [link](example/GPU/HF-Transformers-AutoModels/Model/yuan2) |
|
||||
|
|
|
|||
|
|
@ -0,0 +1,73 @@
|
|||
# DeepSeek-MoE
|
||||
|
||||
In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on DeepSeek-MoE models. For illustration purposes, we utilize the [deepseek-ai/deepseek-moe-16b-chat](https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat) as a reference DeepSeek-MoE 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 DeepSeek-MoE 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 DeepSeek-MoE 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 DeepSeek-MoE 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 'What is AI?'
|
||||
```
|
||||
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 'What is AI？'
|
||||
```
|
||||
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 DeepSeek-MoE model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'deepseek-ai/DeepSeek-MoE-16b-chat'`.
|
||||
- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `What is AI?`.
|
||||
- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
#### 2.4 Sample Output
|
||||
#### [deepseek-ai/DeepSeek-MoE-16b-chat](https://huggingface.co/deepseek-ai/DeepSeek-MoE-16b-chat)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
<｜begin▁of▁sentence｜>[INST] What is AI? [/INST]
|
||||
-------------------- Output --------------------
|
||||
<｜begin▁of▁sentence｜>[INST] What is AI? [/INST]
|
||||
|
||||
Artificial Intelligence (AI) is a field of computer science that aims to create intelligent machines that can perform tasks that typically require human intelligence, such as learning
|
||||
```
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,62 @@
|
|||
#
|
||||
# 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, GenerationConfig
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for DeepSeek-MoE model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="/mnt/disk1/models/deepseek-moe-16b-chat",
|
||||
help='The huggingface repo id for the CodeShell 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')
|
||||
|
||||
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
|
||||
from bigdl.llm.transformers import AutoModelForCausalLM
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, load_in_4bit=True).eval()
|
||||
model.generation_config = GenerationConfig.from_pretrained(model_path)
|
||||
model.generation_config.pad_token_id = model.generation_config.eos_token_id
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path,
|
||||
trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
messages = [
|
||||
{"role": "user", "content": args.prompt}
|
||||
]
|
||||
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
|
||||
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
|
||||
st = time.time()
|
||||
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=args.n_predict)
|
||||
end = time.time()
|
||||
result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
|
||||
print(f'Inference time: {end-st} s')
|
||||
print('-'*20, 'Prompt', '-'*20)
|
||||
print(prompt)
|
||||
print('-'*20, 'Output', '-'*20)
|
||||
print(result)
|
||||
|
|
@ -0,0 +1,63 @@
|
|||
# DeepSeek-MoE
|
||||
In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate DeepSeek-MoE models. For illustration purposes, we utilize the [deepseek-ai/DeepSeek-MoE-16b-chat](https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat) as a reference DeepSeek-MoE 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 deepseek-moe 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
|
||||
```
|
||||
|
||||
### 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 'What is AI?'
|
||||
```
|
||||
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 'What is AI?'
|
||||
```
|
||||
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 DeepSeek-MoE model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'deepseek-ai/deepseek-moe-16b-chat'`.
|
||||
- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `'What is AI?'`.
|
||||
- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
#### 2.4 Sample Output
|
||||
#### [deepseek-ai/deepseek-moe-16b-chat](https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat)
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Prompt --------------------
|
||||
<｜begin▁of▁sentence｜>[INST] What is AI? [/INST]
|
||||
-------------------- Output --------------------
|
||||
<｜begin▁of▁sentence｜>[INST] What is AI? [/INST]
|
||||
|
||||
Artificial Intelligence (AI) is a field of computer science that aims to create intelligent machines that can perform tasks that typically require human intelligence, such as learning
|
||||
```
|
||||
|
||||
|
|
@ -0,0 +1,62 @@
|
|||
#
|
||||
# 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, GenerationConfig
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for DeepSeek-MoE model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="/mnt/disk1/models/deepseek-moe-16b-chat",
|
||||
help='The huggingface repo id for the CodeShell 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')
|
||||
|
||||
args = parser.parse_args()
|
||||
model_path = args.repo_id_or_model_path
|
||||
|
||||
|
||||
from transformers import AutoModelForCausalLM
|
||||
from bigdl.llm import optimize_model
|
||||
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype = torch.bfloat16, device_map = "auto", attn_implementation="eager")
|
||||
model.generation_config = GenerationConfig.from_pretrained(model_path)
|
||||
model.generation_config.pad_token_id = model.generation_config.eos_token_id
|
||||
model = optimize_model(model)
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path,
|
||||
trust_remote_code=True)
|
||||
|
||||
# Generate predicted tokens
|
||||
with torch.inference_mode():
|
||||
messages = [
|
||||
{"role": "user", "content": args.prompt}
|
||||
]
|
||||
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
|
||||
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
|
||||
st = time.time()
|
||||
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=args.n_predict)
|
||||
end = time.time()
|
||||
result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
|
||||
print(f'Inference time: {end-st} s')
|
||||
print('-'*20, 'Prompt', '-'*20)
|
||||
print(prompt)
|
||||
print('-'*20, 'Output', '-'*20)
|
||||
print(result)
|
||||
|
|
@ -861,6 +861,7 @@ def _optimize_post(model, lightweight_bmm=False):
|
|||
# transformers version >= 4.36.0
|
||||
from bigdl.llm.transformers.models.falcon import \
|
||||
falcon_attention_forward_4_36
|
||||
|
||||
convert_forward(model,
|
||||
module.FalconAttention,
|
||||
falcon_attention_forward_4_36
|
||||
|
|
|
|||
Loading…
Reference in a new issue