From 36fbe2144dcc638cb00c4c012bf16ffa8b345dfa Mon Sep 17 00:00:00 2001 From: dingbaorong Date: Thu, 9 Nov 2023 15:29:19 +0800 Subject: [PATCH] Add CPU examples of fuyu (#9393) * add fuyu cpu examples * add gpu example * add comments * add license * remove gpu example * fix inference time --- README.md | 1 + python/llm/README.md | 1 + .../Model/fuyu/README.md | 74 +++++++++++++++++++ .../Model/fuyu/generate.py | 66 +++++++++++++++++ .../CPU/PyTorch-Models/Model/fuyu/README.md | 74 +++++++++++++++++++ .../CPU/PyTorch-Models/Model/fuyu/generate.py | 68 +++++++++++++++++ 6 files changed, 284 insertions(+) create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/README.md create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/generate.py create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/fuyu/README.md create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/fuyu/generate.py diff --git a/README.md b/README.md index 551b0e9b..d9a35bbe 100644 --- a/README.md +++ b/README.md @@ -163,6 +163,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | InternLM-XComposer | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/internlm-xcomposer) | | | WizardCoder-Python | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/wizardcoder-python) | | | CodeShell | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/CodeShell) | | +| Fuyu | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu) | | ***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).*** diff --git a/python/llm/README.md b/python/llm/README.md index a70dcc88..22124728 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -70,6 +70,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | InternLM-XComposer | [link](example/CPU/HF-Transformers-AutoModels/Model/internlm-xcomposer) | | | WizardCoder-Python | [link](example/CPU/HF-Transformers-AutoModels/Model/wizardcoder-python) | | | CodeShell | [link](example/CPU/HF-Transformers-AutoModels/Model/CodeShell) | | +| Fuyu | [link](example/CPU/HF-Transformers-AutoModels/Model/fuyu) | | ### Working with `bigdl-llm` diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/README.md new file mode 100644 index 00000000..410c703c --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/README.md @@ -0,0 +1,74 @@ +# Fuyu +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Fuyu models. For illustration purposes, we utilize the [adept/fuyu-8b](https://huggingface.co/adept/fuyu-8b) as a reference Fuyu 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 an Fuyu 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 transformers==4.35 pillow # additional package required for Fuyu 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 Fuyu 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 --image-path demo.jpg +``` +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-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 --image-path demo.jpg +``` +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 REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Fuyu model (e.g. `adept/fuyu-8b`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'adept/fuyu-8b'`. +- `--prompt PROMPT`: argument defining the prompt to be inferred (with the image for chat). It is default to be `'Generate a coco-style caption.'`. +- `--image-path IMAGE_PATH`: argument defining the input image that the chat will focus on. It is required and should be a local path (not url). +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `512`. + + +#### 2.4 Sample Output +#### [adept/fuyu-8b](https://huggingface.co/adept/fuyu-8b) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +Generate a coco-style caption. +-------------------- Output -------------------- +An orange bus parked on the side of a road. +``` + +The sample input image is (which is fetched from [COCO dataset](https://cocodataset.org/#explore?id=178242)): + +[demo.jpg](https://cocodataset.org/#explore?id=178242) + + \ No newline at end of file diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/generate.py new file mode 100644 index 00000000..7fe83502 --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu/generate.py @@ -0,0 +1,66 @@ +# +# 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. +# + +from transformers import FuyuProcessor +import torch +import argparse +import time +from PIL import Image +from bigdl.llm.transformers import AutoModelForCausalLM + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Fuyu model') + parser.add_argument('--repo-id-or-model-path', type=str, default="adept/fuyu-8b", + help='The huggingface repo id for the Fuyu model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="Generate a coco-style caption.", + help='Prompt to infer') + parser.add_argument('--image-path', type=str, required=True, + help='Image path for the input image that the chat will focus on') + parser.add_argument('--n-predict', type=int, default=512, help='Max tokens to predict') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + prompt = args.prompt + image = Image.open(args.image_path) + + # Load model + # For successful BigDL-LLM optimization on Fuyu, skip the 'vision_embed_tokens' module during optimization + model = AutoModelForCausalLM.from_pretrained(model_path, device_map='cpu', + load_in_4bit = True, + trust_remote_code=True, + modules_to_not_convert=['vision_embed_tokens']) + + # Load processor + processor = FuyuProcessor.from_pretrained(model_path) + + # Generate predicted tokens + with torch.inference_mode(): + inputs = processor(text=prompt, images=image, return_tensors="pt") + st = time.time() + generation_outputs = model.generate(**inputs, + max_new_tokens=args.n_predict) + end = time.time() + outputs = processor.batch_decode(generation_outputs[:, -args.n_predict:], skip_special_tokens=True) + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + for output in outputs: + # '\x04' is the "beginning of answer" token + # See https://huggingface.co/adept/fuyu-8b#how-to-use + answer = output.split('\x04 ', 1)[1] if '\x04' in output else '' + print(answer) diff --git a/python/llm/example/CPU/PyTorch-Models/Model/fuyu/README.md b/python/llm/example/CPU/PyTorch-Models/Model/fuyu/README.md new file mode 100644 index 00000000..410c703c --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/fuyu/README.md @@ -0,0 +1,74 @@ +# Fuyu +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Fuyu models. For illustration purposes, we utilize the [adept/fuyu-8b](https://huggingface.co/adept/fuyu-8b) as a reference Fuyu 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 an Fuyu 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 transformers==4.35 pillow # additional package required for Fuyu 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 Fuyu 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 --image-path demo.jpg +``` +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-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 --image-path demo.jpg +``` +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 REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Fuyu model (e.g. `adept/fuyu-8b`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'adept/fuyu-8b'`. +- `--prompt PROMPT`: argument defining the prompt to be inferred (with the image for chat). It is default to be `'Generate a coco-style caption.'`. +- `--image-path IMAGE_PATH`: argument defining the input image that the chat will focus on. It is required and should be a local path (not url). +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `512`. + + +#### 2.4 Sample Output +#### [adept/fuyu-8b](https://huggingface.co/adept/fuyu-8b) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +Generate a coco-style caption. +-------------------- Output -------------------- +An orange bus parked on the side of a road. +``` + +The sample input image is (which is fetched from [COCO dataset](https://cocodataset.org/#explore?id=178242)): + +[demo.jpg](https://cocodataset.org/#explore?id=178242) + + \ No newline at end of file diff --git a/python/llm/example/CPU/PyTorch-Models/Model/fuyu/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/fuyu/generate.py new file mode 100644 index 00000000..234e4741 --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/fuyu/generate.py @@ -0,0 +1,68 @@ +# +# 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. +# + +from transformers import AutoModelForCausalLM, FuyuProcessor +import torch +import argparse +import time +from PIL import Image +from bigdl.llm import optimize_model + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Fuyu model') + parser.add_argument('--repo-id-or-model-path', type=str, default="adept/fuyu-8b", + help='The huggingface repo id for the Fuyu model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="Generate a coco-style caption.", + help='Prompt to infer') + parser.add_argument('--image-path', type=str, required=True, + help='Image path for the input image that the chat will focus on') + parser.add_argument('--n-predict', type=int, default=512, help='Max tokens to predict') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + prompt = args.prompt + image = Image.open(args.image_path) + + # Load model + model = AutoModelForCausalLM.from_pretrained(model_path, device_map='cpu', trust_remote_code=True) + + # With only one line to enable BigDL-LLM optimization on model + # For successful BigDL-LLM optimization on Fuyu, skip the 'vision_embed_tokens' module during optimization + model = optimize_model(model, + low_bit='sym_int4', + modules_to_not_convert=['vision_embed_tokens']) + + # Load processor + processor = FuyuProcessor.from_pretrained(model_path) + + # Generate predicted tokens + with torch.inference_mode(): + inputs = processor(text=prompt, images=image, return_tensors="pt") + st = time.time() + generation_outputs = model.generate(**inputs, + max_new_tokens=args.n_predict) + end = time.time() + outputs = processor.batch_decode(generation_outputs[:, -args.n_predict:], skip_special_tokens=True) + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + for output in outputs: + # '\x04' is the "beginning of answer" token + # See https://huggingface.co/adept/fuyu-8b#how-to-use + answer = output.split('\x04 ', 1)[1] if '\x04' in output else '' + print(answer)