Skip to content

Commit

Permalink
Add awq load support (intel-analytics#9453)
Browse files Browse the repository at this point in the history
* Support directly loading GPTQ models from huggingface

* fix style

* fix tests

* change example structure

* address comments

* fix style

* init

* address comments

* add examples

* fix style

* fix style

* fix style

* fix style

* update

* remove

* meet comments

* fix style

---------

Co-authored-by: Yang Wang <[email protected]>
  • Loading branch information
cyita and yangw1234 authored Nov 16, 2023
1 parent 3d4f147 commit 7d7f1f9
Show file tree
Hide file tree
Showing 11 changed files with 1,090 additions and 19 deletions.
Original file line number Diff line number Diff line change
@@ -0,0 +1,71 @@
# AWQ
This example shows how to directly run 4-bit AWQ models using BigDL-LLM on Intel CPU. For illustration purposes, we utilize the ["TheBloke/Llama-2-7B-Chat-AWQ"](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ) as a reference.

## 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 Llama2 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 autoawq==0.1.6 --no-deps
pip install bigdl-llm[all] # install bigdl-llm with 'all' option
pip install transformers==4.35.0
pip install accelerate==0.24.1
```

### 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 Llama2-awq model (e.g. `TheBloke/Llama-2-7B-Chat-AWQ`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'TheBloke/Llama-2-7B-Chat-AWQ'`.
- `--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`.

> **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 Llama2 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-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
```

#### 2.3 Sample Output
#### ["TheBloke/Llama-2-7B-Chat-AWQ"](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
### HUMAN:
What is AI?
### RESPONSE:
-------------------- Output --------------------
### HUMAN:
What is AI?
### RESPONSE:
Artificial intelligence (AI) is the ability of machines to perform tasks that typically require human intelligence, such as learning, problem-solving, decision
```
Original file line number Diff line number Diff line change
@@ -0,0 +1,71 @@
#
# 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 LlamaTokenizer

# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://huggingface.co/georgesung/llama2_7b_chat_uncensored#prompt-style
LLAMA2_PROMPT_FORMAT = """### HUMAN:
{prompt}
### RESPONSE:
"""

if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="TheBloke/Llama-2-7B-Chat-AWQ",
help='The huggingface repo id'
', 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
model = AutoModelForCausalLM.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True)

# Load tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Generate predicted tokens
with torch.inference_mode():
prompt = LLAMA2_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)
Original file line number Diff line number Diff line change
@@ -0,0 +1,65 @@
# AWQ
This example shows how to directly run 4-bit AWQ models using BigDL-LLM on Intel GPU. For illustration purposes, we utilize the ["TheBloke/Llama-2-7B-Chat-AWQ"](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ) as a reference.

## 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 Llama2 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 --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
pip install transformers==4.35.0
pip install autoawq==0.1.6 --no-deps
pip install accelerate==0.24.1
```

### 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 Llama2-awq model (e.g. `TheBloke/Llama-2-7B-Chat-AWQ`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'TheBloke/Llama-2-7B-Chat-AWQ'`.
- `--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`.

> **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 Llama2 model based on the capabilities of your machine.
#### 2.3 Sample Output
#### ["TheBloke/Llama-2-7B-Chat-AWQ"](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
### HUMAN:
What is AI?
### RESPONSE:
-------------------- Output --------------------
### HUMAN:
What is AI?
### RESPONSE:
Artificial intelligence (AI) is the ability of machines to perform tasks that typically require human intelligence, such as learning, problem-solving, decision
```
Original file line number Diff line number Diff line change
@@ -0,0 +1,71 @@
#
# 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 intel_extension_for_pytorch as ipex
from bigdl.llm.transformers import AutoModelForCausalLM
from transformers import LlamaTokenizer

# you could tune the prompt based on your own model,
# here the prompt tuning refers to https://huggingface.co/georgesung/llama2_7b_chat_uncensored#prompt-style
LLAMA2_PROMPT_FORMAT = """### HUMAN:
{prompt}
### RESPONSE:
"""

if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
parser.add_argument('--repo-id-or-model-path', type=str, default="TheBloke/Llama-2-7B-Chat-AWQ",
help='The huggingface repo id'
', 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
model = AutoModelForCausalLM.from_pretrained(model_path,
load_in_4bit=True,
trust_remote_code=True,).to("xpu")

# Load tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Generate predicted tokens
with torch.inference_mode():
prompt = LLAMA2_PROMPT_FORMAT.format(prompt=args.prompt)
input_ids = tokenizer.encode(prompt, return_tensors="pt").to("xpu")
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)
21 changes: 21 additions & 0 deletions python/llm/src/bigdl/llm/transformers/awq/__init__.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,21 @@
#
# 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.
#

# This would makes sure Python is aware there is more than one sub-package within bigdl,
# physically located elsewhere.
# Otherwise there would be module not found error in non-pip's setting as Python would
# only search the first bigdl package and end up finding only one sub-package.

54 changes: 54 additions & 0 deletions python/llm/src/bigdl/llm/transformers/awq/act.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,54 @@
#
# 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.
#
# ===========================================================================
#
# This file is copied from
# https://github.com/casper-hansen/AutoAWQ/blob/main/awq/modules/act.py
#
# MIT License
#
# Copyright (c) 2023 MIT HAN Lab
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
#

import torch.nn as nn


class ScaledActivation(nn.Module):
def __init__(self, module, scales):
super().__init__()
self.act = module
self.scales = nn.Parameter(scales.data)

def forward(self, x):
return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
Loading

0 comments on commit 7d7f1f9

Please sign in to comment.