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

Update llm README.md (#8431)

Browse source
This commit is contained in:
Jason Dai 2023-06-30 19:41:17 +08:00 committed by GitHub
parent 2fd751de7a
commit 2da21163f8
1 changed files with 109 additions and 109 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

202
python/llm/README.md
View file

@ -1,148 +1,148 @@
# BigDL LLM ## BigDL-LLM
`bigdl-llm` is an SDK for large language model (LLM). It helps users develop AI applications that contains LLM on Intel XPU by using less computing and memory resources.`bigdl-llm` utilize a highly optimized GGML on Intel XPU.
Users could use `bigdl-llm` to **`bigdl-llm`** is a library for running ***LLM*** (language language model) on your Intel ***laptop*** using INT4 with very low latency.
- Convert their model to lower precision
- Use command line tool like `llama.cpp` to run the model inference
- Use transformers like API to run the model inference
- Integrate the model in `langchain` pipeline
Currently `bigdl-llm` has supported *(It is built on top of the excellent work of [llama.cpp](https://github.com/ggerganov/llama.cpp), [gptq](https://github.com/IST-DASLab/gptq), [ggml](https://github.com/ggerganov/ggml), [llama-cpp-python](https://github.com/abetlen/llama-cpp-python), [gptq_for_llama](https://github.com/qwopqwop200/GPTQ-for-LLaMa), [bitsandbytes](https://github.com/TimDettmers/bitsandbytes), [redpajama.cpp](https://github.com/togethercomputer/redpajama.cpp), [gptneox.cpp](https://github.com/byroneverson/gptneox.cpp), [bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp/), etc.)*
- Precision: INT4
- Model Family: llama, gptneox, bloom, starcoder
- Platform: Ubuntu 20.04 or later, CentOS 7 or later, Windows 10/11
- Device: CPU
- Python: 3.9 (recommended) or later
## Installation ### Demos
BigDL-LLM is a self-contained SDK library for model loading and inferencing. Users could directly See the ***optimized performance*** of `phoenix-inst-chat-7b`, `vicuna-13b-v1.1`, and `starcoder-15b` models on a 12th Gen Intel Core CPU below.
```bash
pip install --pre --upgrade bigdl-llm <p align="center">
``` <img src="https://github.com/bigdl-project/bigdl-project.github.io/blob/master/assets/llm-7b.gif" width='33%' /> <img src="https://github.com/bigdl-project/bigdl-project.github.io/blob/master/assets/llm-13b.gif" width='33%' /> <img src="https://github.com/bigdl-project/bigdl-project.github.io/blob/master/assets/llm-15b5.gif" width='33%' />
While model conversion procedure will rely on some 3rd party libraries. Add `[all]` option for installation to prepare environment. <img src="https://github.com/bigdl-project/bigdl-project.github.io/blob/master/assets/llm-models.png" width='85%'/>
</p>
### Working with `bigdl-llm`
#### Install
You may install **`bigdl-llm`** as follows:
```bash ```bash
pip install --pre --upgrade bigdl-llm[all] pip install --pre --upgrade bigdl-llm[all]
``` ```
#### Download Model
## Usage You may download any PyTorch model in Hugging Face *Transformers* format (including *FP16* or *FP32* or *GPTQ-4bit*).
A standard procedure for using `bigdl-llm` contains 3 steps:
1. Download model from huggingface hub #### Run Model
2. Convert model from huggingface format to GGML format
3. Inference using `llm-cli`, transformers like API, or `langchain`.
### Convert your model You may run the models using **`bigdl-llm`** through one of the following APIs:
A python function and a command line tool `llm-convert` is provided to transform the model from huggingface format to GGML format. 1. [CLI (command line interface) Tool](#cli-tool)
2. [Hugging Face `transformer`-style API](#hugging-face-transformers-style-api)
3. [LangChain API](#langchain-api)
4. [`llama-cpp-python`-style API](#llama-cpp-python-style-api)
#### CLI Tool
Currently `bigdl-llm` CLI supports *LLaMA* (e.g., *vicuna*), *GPT-NeoX* (e.g., *redpajama*), *BLOOM* (e.g., *pheonix*) and *GPT2* (e.g., *starcoder*) model architecture; for other models, you may use the `transformer`-style or LangChain APIs.
- ##### Convert model
You may convert the downloaded model into native INT4 format using `llm-convert`.
Here is an example to use `llm-convert` command line tool.
```bash ```bash
# pth model #convert PyTorch (fp16 or fp32) model;
llm-convert "/path/to/llama-7b-hf/" --model-format pth --outfile "/path/to/llama-7b-int4/" --model-family "llama" #llama/bloom/gptneox/starcoder model family is currently supported
# gptq model lm-convert "/path/to/model/" --model-format pth --model-family "bloom" --outfile "/path/to/output/"
llm-convert "/path/to/vicuna-13B-1.1-GPTQ-4bit-128g/" --model-format gptq --outfile "/path/to/vicuna-13B-int4/" --model-family "llama"
```
> An example GPTQ model can be found [here](https://huggingface.co/TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g/tree/main)
Here is an example to use `llm_convert` python API. #convert GPTQ-4bit model
```bash #only llama model family is currently supported
from bigdl.llm import llm_convert llm-convert "/path/to/model/" --model-format gptq --model-family "llama" --outfile "/path/to/output/"
# pth model
llm_convert(model="/path/to/llama-7b-hf/",
outfile="/path/to/llama-7b-int4/",
model_format="pth",
model_family="llama")
# gptq model
llm_convert(model="/path/to/vicuna-13B-1.1-GPTQ-4bit-128g/",
outfile="/path/to/vicuna-13B-int4/",
model_format="gptq",
model_family="llama")
``` ```
### Inferencing - ##### Run model
#### llm-cli command line You may run the converted model using `llm-cli` (*built on top of `main.cpp` in [llama.cpp](https://github.com/ggerganov/llama.cpp)*)
llm-cli is a command-line interface tool that follows the interface as the main program in `llama.cpp`.
```bash ```bash
#help
#llama/bloom/gptneox/starcoder model family is currently supported
llm-cli -x gptneox -h
#text completion #text completion
llm-cli -t 16 -x llama -m "/path/to/llama-7b-int4/bigdl-llm-xxx.bin" -p 'Once upon a time,' #llama/bloom/gptneox/starcoder model family is currently supported
llm-cli -t 16 -x gptneox -m "/path/to/output/model.bin" -p 'Once upon a time,'
# chatting
llm-cli -t 16 -x llama -m "/path/to/llama-7b-int4/bigdl-llm-xxx.bin" -i --color
# help information
llm-cli -x llama -h
``` ```
#### Transformers like API #### Hugging Face `transformers`-style API
You can also load the converted model using `BigdlForCausalLM` with a transformer like API, You may run the models using `transformers`-style API in `bigdl-llm`
- ##### Using native INT4 format
You may convert Hugging Face *Transformers* models into native INT4 format for maximum performance as follows.
*(Currently only llama/bloom/gptneox/starcoder model family is supported; for other models, you may use the [Hugging Face `transformers` INT4 format](#using-hugging-face-transformers-int4-format)).*
```python ```python
#convert the model
from bigdl.llm import llm_convert
bigdl_llm_path = llm_convert(model='/path/to/model/',
outfile='/path/to/output/', outtype='int4', model_family="llama")
#load the converted model
from bigdl.llm.transformers import BigdlForCausalLM from bigdl.llm.transformers import BigdlForCausalLM
llm = BigdlForCausalLM.from_pretrained("/path/to/llama-7b-int4/bigdl-llm-xxx.bin", llm = BigdlForCausalLM.from_pretrained("/path/to/output/model.bin",...)
model_family="llama")
prompt="What is AI?" #run the converted model
``` input_ids = llm.tokenize(prompt)
and simply do inference end-to-end like output_ids = llm.generate(input_ids, ...)
```python output = llm.batch_decode(output_ids)
output = llm(prompt, max_tokens=32)
```
If you need to seperate the tokenization and generation, you can also do inference like
```python
tokens_id = llm.tokenize(prompt)
output_tokens_id = llm.generate(tokens_id, max_new_tokens=32)
output = llm.batch_decode(output_tokens_id)
``` ```
- ##### Using Hugging Face `transformers` INT4 format
Alternatively, you can load huggingface model directly using `AutoModelForCausalLM.from_pretrained`. You may apply INT4 optimizations to any Hugging Face *Transformers* models as follows.
```python ```python
#load Hugging Face Transformers model with INT4 optimizations
from bigdl.llm.transformers import AutoModelForCausalLM from bigdl.llm.transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained('/path/to/model/', load_in_4bit=True)
# option 1: load huggingface checkpoint #run the optimized model
llm = AutoModelForCausalLM.from_pretrained("/path/to/llama-7b-hf/", from transformers import AutoTokenizer
model_family="llama") tokenizer = AutoTokenizer.from_pretrained(model_path)
input_ids = tokenizer.encode(input_str, ...)
# option 2: load from huggingface hub repo output_ids = model.generate(input_ids, ...)
llm = AutoModelForCausalLM.from_pretrained("decapoda-research/llama-7b-hf", output = tokenizer.batch_decode(output_ids)
model_family="llama")
``` ```
You can then use the the model the same way as you use transformers. #### LangChain API
You may convert Hugging Face *Transformers* models into *native INT4* format (currently only *llama*/*bloom*/*gptneox*/*starcoder* model family is supported), and then run the converted models using the LangChain API in `bigdl-llm` as follows.
```python ```python
# Use transformers tokenizer from bigdl.llm.langchain.llms import BigdlLLM
tokenizer = AutoTokenizer.from_pretrained(model_ckpt) from bigdl.llm.langchain.embeddings import BigdlLLMEmbeddings
tokens = tokenizer("what is ai").input_ids from langchain.chains.question_answering import load_qa_chain
tokens_id = llm.generate(tokens, max_new_tokens=32)
tokenizer.batch_decode(tokens_id) embeddings = BigdlLLMEmbeddings(model_path='/path/to/converted/model.bin',
model_family="llama",...)
bigdl_llm = BigdlLLM(model_path='/path/to/converted/model.bin',
model_family="llama",...)
doc_chain = load_qa_chain(bigdl_llm, ...)
doc_chain.run(...)
``` ```
#### llama-cpp-python like API #### `llama-cpp-python`-style API
`llama-cpp-python` has become a popular pybinding for `llama.cpp` program. Some users may be familiar with this API so `bigdl-llm` reserve this API and extend it to other model families (e.g., gptneox, bloom)
You may also run the converted models using the `llama-cpp-python`-style API in `bigdl-llm` as follows.
```python ```python
from bigdl.llm.models import Llama, Bloom, Gptneox, Starcoder from bigdl.llm.models import Llama, Bloom, Gptneox
llm = Llama("/path/to/llama-7b-int4/bigdl-llm-xxx.bin", n_threads=4) llm = Bloom("/path/to/converted/model.bin", n_threads=4)
result = llm("what is ai") result = llm("what is ai")
``` ```
#### langchain integration ### `bigdl-llm` Dependence
TODO The native code/lib in `bigdl-llm` has been built using the following tools; in particular, lower `LIBC` version on your Linux system may be incompatible with `bigdl-llm`.
## Examples
We prepared several examples in https://github.com/intel-analytics/BigDL/tree/main/python/llm/example
## Dynamic library BOM
To avoid difficaulties during the installtion. `bigdl-llm` release the C implementation by dynamic library or executive file. The compilation details are stated below. **These information is only for reference, no compilation procedure is needed for our users.** `GLIBC` version may affect the compatibility.
| Model family | Platform | Compiler | GLIBC | | Model family | Platform | Compiler | GLIBC |
| ------------ | -------- | ------------------ | ----- | | ------------ | -------- | ------------------ | ----- |
| llama | Linux | GCC 9.4.0 | 2.17 | | llama | Linux | GCC 9.3.1 | 2.17 |
| llama | Windows | MSVC 19.36.32532.0 | | | llama | Windows | MSVC 19.36.32532.0 | |
| gptneox | Linux | GCC 9.4.0 | 2.17 | | gptneox | Linux | GCC 9.3.1 | 2.17 |
| gptneox | Windows | MSVC 19.36.32532.0 | | | gptneox | Windows | MSVC 19.36.32532.0 | |
| bloom | Linux | GCC 9.4.0 | 2.31 | | bloom | Linux | GCC 9.4.0 | 2.29 |
| bloom | Windows | MSVC 19.36.32532.0 | | | bloom | Windows | MSVC 19.36.32532.0 | |
| starcoder | Linux | GCC 9.4.0 | 2.31 | | starcoder | Linux | GCC 9.4.0 | 2.29 |
| starcoder | Windows | MSVC 19.36.32532.0 | | | starcoder | Windows | MSVC 19.36.32532.0 | |