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

LLM: support Mistral AWQ models (#9520)

Browse source
This commit is contained in:
binbin Deng 2023-11-24 16:20:22 +08:00 committed by GitHub
parent 914a5a5a27
commit 6bec0faea5
5 changed files with 34 additions and 22 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
python/llm/example/CPU/HF-Transformers-AutoModels/Advanced-Quantizations/AWQ/README.md
View file

@ -1,11 +1,16 @@
# 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.
This example shows how to directly run 4-bit AWQ models using BigDL-LLM on Intel CPU.
## 0. Requirements
## Verified Models
- [Llama-2-7B-Chat-AWQ](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ)
- [Mistral-7B-Instruct-v0.1-AWQ](https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-AWQ)
- [Mistral-7B-v0.1-AWQ](https://huggingface.co/TheBloke/Mistral-7B-v0.1-AWQ)
## 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.
In the example [generate.py](./generate.py), we show a basic use case for a AWQ 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
@ -13,7 +18,7 @@ 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 --pre --upgrade bigdl-llm[all] # install bigdl-llm with 'all' option
pip install transformers==4.35.0
pip install accelerate==0.24.1
```
@ -24,13 +29,13 @@ python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROM
```
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'`.
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the AWQ model (e.g. `TheBloke/Llama-2-7B-Chat-AWQ`, `TheBloke/Mistral-7B-Instruct-v0.1-AWQ`, `TheBloke/Mistral-7B-v0.1-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.
> Please select the appropriate size of the 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:
@ -52,7 +57,7 @@ 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)
#### [TheBloke/Llama-2-7B-Chat-AWQ](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ)
```log
Inference time: xxxx s
-------------------- Prompt --------------------
@ -68,4 +73,4 @@ 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
```
```

10
python/llm/example/CPU/HF-Transformers-AutoModels/Advanced-Quantizations/AWQ/generate.py
View file

@ -19,18 +19,18 @@ import time
import argparse
from bigdl.llm.transformers import AutoModelForCausalLM
from transformers import LlamaTokenizer
from transformers import AutoTokenizer
# 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_FORMAT = """### HUMAN:
{prompt}
### RESPONSE:
"""
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for AWQ 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')
@ -49,11 +49,11 @@ if __name__ == '__main__':
trust_remote_code=True)
# Load tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = LLAMA2_PROMPT_FORMAT.format(prompt=args.prompt)
prompt = 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

13
python/llm/example/GPU/HF-Transformers-AutoModels/Advanced-Quantizations/AWQ/README.md
View file

@ -1,11 +1,16 @@
# 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.
This example shows how to directly run 4-bit AWQ models using BigDL-LLM on Intel GPU.
## 0. Requirements
## Verified Models
- [Llama-2-7B-Chat-AWQ](https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ)
- [Mistral-7B-Instruct-v0.1-AWQ](https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-AWQ)
- [Mistral-7B-v0.1-AWQ](https://huggingface.co/TheBloke/Mistral-7B-v0.1-AWQ)
## 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.
In the example [generate.py](./generate.py), we show a basic use case for a AWQ 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
@ -37,7 +42,7 @@ python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROM
```
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'`.
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the AWQ model (e.g. `TheBloke/Llama-2-7B-Chat-AWQ`, `TheBloke/Mistral-7B-Instruct-v0.1-AWQ`, `TheBloke/Mistral-7B-v0.1-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`.

10
python/llm/example/GPU/HF-Transformers-AutoModels/Advanced-Quantizations/AWQ/generate.py
View file

@ -19,18 +19,18 @@ import time
import argparse
import intel_extension_for_pytorch as ipex
from bigdl.llm.transformers import AutoModelForCausalLM
from transformers import LlamaTokenizer
from transformers import AutoTokenizer
# 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_FORMAT = """### HUMAN:
{prompt}
### RESPONSE:
"""
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for AWQ 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')
@ -49,11 +49,11 @@ if __name__ == '__main__':
trust_remote_code=True,).to("xpu")
# Load tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
# Generate predicted tokens
with torch.inference_mode():
prompt = LLAMA2_PROMPT_FORMAT.format(prompt=args.prompt)
prompt = 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

2
python/llm/src/bigdl/llm/transformers/awq/awq.py
View file

@ -131,6 +131,8 @@ def get_blocks(model):
layers = model.transformer.h
elif "neox" in str(model.__class__).lower():
layers = model.gpt_neox.layers
elif "mistral" in str(model.__class__).lower():
layers = model.model.layers
else:
invalidInputError(False, f"Model type {type(model)} isn't supported.")
return layers