From 3f24202e4c6368446a063be2d4df8c50b36dd5df Mon Sep 17 00:00:00 2001 From: binbin Deng <108676127+plusbang@users.noreply.github.com> Date: Tue, 25 Jul 2023 09:21:12 +0800 Subject: [PATCH] [LLM] Add more transformers int4 example (Llama 2) (#8602) --- python/llm/README.md | 1 + .../transformers/transformers_int4/README.md | 1 + .../transformers_int4/llama2/README.md | 86 +++++++++++++++++++ .../transformers_int4/llama2/generate.py | 71 +++++++++++++++ 4 files changed, 159 insertions(+) create mode 100644 python/llm/example/transformers/transformers_int4/llama2/README.md create mode 100644 python/llm/example/transformers/transformers_int4/llama2/generate.py diff --git a/python/llm/README.md b/python/llm/README.md index 4216fbe8..108e0683 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -17,6 +17,7 @@ We may use any Hugging Face Transfomer models on `bigdl-llm`, and the following | Model | Example | |-----------|----------------------------------------------------------| | LLaMA *(such as Vicuna, Guanaco, Koala, Baize, WizardLM, etc.)* | [link1](example/transformers/native_int4), [link2](example/transformers/transformers_int4/vicuna) | +| LLaMA 2 | [link](example/transformers/transformers_int4/llama2) | | MPT | [link](example/transformers/transformers_int4/mpt) | | Falcon | [link](example/transformers/transformers_int4/falcon) | | ChatGLM | [link](example/transformers/transformers_int4/chatglm) | diff --git a/python/llm/example/transformers/transformers_int4/README.md b/python/llm/example/transformers/transformers_int4/README.md index 9a2379ed..4224caeb 100644 --- a/python/llm/example/transformers/transformers_int4/README.md +++ b/python/llm/example/transformers/transformers_int4/README.md @@ -5,6 +5,7 @@ You can use BigDL-LLM to run any Huggingface Transformer models with INT4 optimi | Model | Example | |-----------|----------------------------------------------------------| | LLaMA | [link](vicuna) | +| LLaMA 2 | [link](llama2) | | MPT | [link](mpt) | | Falcon | [link](falcon) | | ChatGLM | [link](chatglm) | diff --git a/python/llm/example/transformers/transformers_int4/llama2/README.md b/python/llm/example/transformers/transformers_int4/llama2/README.md new file mode 100644 index 00000000..dc61777c --- /dev/null +++ b/python/llm/example/transformers/transformers_int4/llama2/README.md @@ -0,0 +1,86 @@ +# Llama2 +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Llama2 models. For illustration purposes, we utilize the [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) and [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf) as reference Llama2 models. + +## 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 bigdl-llm[all] # install bigdl-llm with 'all' option +``` + +### 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 model (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'meta-llama/Llama-2-7b-chat-hf'`. +- `--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-nano-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 +#### [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +### HUMAN: +What is AI? + +### RESPONSE: + +-------------------- Output -------------------- +### HUMAN: +What is AI? + +### RESPONSE: + +AI is a term used to describe the development of computer systems that can perform tasks that typically require human intelligence, such as understanding natural language, recognizing images +``` + +#### [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +### HUMAN: +What is AI? + +### RESPONSE: + +-------------------- Output -------------------- +### HUMAN: +What is AI? + +### RESPONSE: + +AI, or artificial intelligence, refers to the ability of machines to perform tasks that would typically require human intelligence, such as learning, problem-solving, +``` diff --git a/python/llm/example/transformers/transformers_int4/llama2/generate.py b/python/llm/example/transformers/transformers_int4/llama2/generate.py new file mode 100644 index 00000000..6353d848 --- /dev/null +++ b/python/llm/example/transformers/transformers_int4/llama2/generate.py @@ -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="meta-llama/Llama-2-7b-chat-hf", + help='The huggingface repo id for the Llama2 (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) 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 + 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)