From 5a2ce421af20ddd5d619477eb352b45db269ab94 Mon Sep 17 00:00:00 2001 From: dingbaorong Date: Tue, 24 Oct 2023 15:24:01 +0800 Subject: [PATCH] add cpu and gpu examples of flan-t5 (#9171) * add cpu and gpu examples of flan-t5 * address yuwen's comments * Add explanation why we add modules to not convert * Refine prompt and add a translation example * Add a empty line at the end of files * add examples of flan-t5 using optimize_mdoel api * address bin's comments * address binbin's comments * add flan-t5 in readme --- README.md | 1 + python/llm/README.md | 2 +- .../Model/README.md | 1 + .../Model/flan-t5/README.md | 66 +++++++++++++++ .../Model/flan-t5/generate.py | 73 ++++++++++++++++ .../CPU/PyTorch-Models/Model/README.md | 1 + .../PyTorch-Models/Model/flan-t5/README.md | 66 +++++++++++++++ .../PyTorch-Models/Model/flan-t5/generate.py | 65 +++++++++++++++ .../Model/README.md | 1 + .../Model/flan-t5/README.md | 55 ++++++++++++ .../Model/flan-t5/generate.py | 83 +++++++++++++++++++ .../GPU/PyTorch-Models/Model/README.md | 1 + .../PyTorch-Models/Model/flan-t5/README.md | 55 ++++++++++++ .../PyTorch-Models/Model/flan-t5/generate.py | 78 +++++++++++++++++ 14 files changed, 547 insertions(+), 1 deletion(-) create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/README.md create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/flan-t5/README.md create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/flan-t5/generate.py create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/README.md create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/flan-t5/README.md create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/flan-t5/generate.py diff --git a/README.md b/README.md index 1b384ebe..ce84f188 100644 --- a/README.md +++ b/README.md @@ -152,6 +152,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | MOSS | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/moss) | | | Whisper | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/whisper) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/whisper) | | Phi-1_5 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5) | +| Flan-t5 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5) | ***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 7e1c2233..4fcab1e4 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -59,7 +59,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | MOSS | [link](example/CPU/HF-Transformers-AutoModels/Model/moss) | | | Whisper | [link](example/CPU/HF-Transformers-AutoModels/Model/whisper) | [link](example/GPU/HF-Transformers-AutoModels/Model/whisper) | | Phi-1_5 | [link](example/CPU/HF-Transformers-AutoModels/Model/phi-1_5) | [link](example/GPU/HF-Transformers-AutoModels/Model/phi-1_5) | - +| Flan-t5 | [link](example/CPU/HF-Transformers-AutoModels/Model/flan-t5) | [link](example/GPU/HF-Transformers-AutoModels/Model/flan-t5) | ### Working with `bigdl-llm` diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/README.md index 55d4c41b..92cb9103 100644 --- a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/README.md +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/README.md @@ -24,6 +24,7 @@ You can use BigDL-LLM to run any Huggingface Transformer models with INT4 optimi | Aquila | [link](aquila) | | Replit | [link](replit) | | Mistral | [link](mistral) | +| Flan-t5 | [link](flan-t5) | ## Recommended Requirements To run the examples, we recommend using Intel® Xeon® processors (server), or >= 12th Gen Intel® Core™ processor (client). diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/README.md new file mode 100644 index 00000000..0b6ab57a --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/README.md @@ -0,0 +1,66 @@ +# Flan-t5 + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Flan-t5 models. For illustration purposes, we utilize the [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) as a reference Flan-t5 model. + +## 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 Flan-t5 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 +``` + +### 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 Flan-t5 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 --prompt 'Translate to German: My name is Arthur' +``` +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 --prompt 'Translate to German: My name is Arthur' +``` +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 Flan-t5 model (e.g. `google/flan-t5-xxl`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/flan-t5-xxl'`. +- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Translate to German: My name is Arthur'`. +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`. + + +#### 2.4 Sample Output +#### [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +<|User|>:Translate to German: My name is Arthur +-------------------- Output -------------------- +Ich bin Arthur. +``` diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py new file mode 100644 index 00000000..58d5e446 --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py @@ -0,0 +1,73 @@ +# +# 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 numpy as np + +from bigdl.llm.transformers import AutoModelForSeq2SeqLM +from transformers import AutoTokenizer + +# you could tune the prompt based on your own model, +FLAN_T5_PROMPT_FORMAT = "<|User|>:{prompt}" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for flan-t5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="google/flan-t5-xxl", + help='The huggingface repo id for the flan-t5 model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="Translate to German: My name is Arthur", + 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. + # "wo" module is not converted due to some issues of T5 model + # (https://github.com/huggingface/transformers/issues/20287), + # "lm_head" module is not converted to generate outputs with better quality + model = AutoModelForSeq2SeqLM.from_pretrained(model_path, + load_in_4bit=True, + trust_remote_code=True, + modules_to_not_convert=["wo", "lm_head"]) + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = FLAN_T5_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) + output_str = output_str.split("")[0] + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + print(output_str) diff --git a/python/llm/example/CPU/PyTorch-Models/Model/README.md b/python/llm/example/CPU/PyTorch-Models/Model/README.md index 268534df..3cee8c45 100644 --- a/python/llm/example/CPU/PyTorch-Models/Model/README.md +++ b/python/llm/example/CPU/PyTorch-Models/Model/README.md @@ -10,6 +10,7 @@ You can use `optimize_model` API to accelerate general PyTorch models on Intel s | BERT | [link](bert) | | Bark | [link](bark) | | Mistral | [link](mistral) | +| Flan-t5 | [link](flan-t5) | ## Recommended Requirements To run the examples, we recommend using Intel® Xeon® processors (server), or >= 12th Gen Intel® Core™ processor (client). diff --git a/python/llm/example/CPU/PyTorch-Models/Model/flan-t5/README.md b/python/llm/example/CPU/PyTorch-Models/Model/flan-t5/README.md new file mode 100644 index 00000000..0b6ab57a --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/flan-t5/README.md @@ -0,0 +1,66 @@ +# Flan-t5 + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Flan-t5 models. For illustration purposes, we utilize the [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) as a reference Flan-t5 model. + +## 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 Flan-t5 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 +``` + +### 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 Flan-t5 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 --prompt 'Translate to German: My name is Arthur' +``` +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 --prompt 'Translate to German: My name is Arthur' +``` +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 Flan-t5 model (e.g. `google/flan-t5-xxl`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/flan-t5-xxl'`. +- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Translate to German: My name is Arthur'`. +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`. + + +#### 2.4 Sample Output +#### [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +<|User|>:Translate to German: My name is Arthur +-------------------- Output -------------------- +Ich bin Arthur. +``` diff --git a/python/llm/example/CPU/PyTorch-Models/Model/flan-t5/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/flan-t5/generate.py new file mode 100644 index 00000000..51ba2500 --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/flan-t5/generate.py @@ -0,0 +1,65 @@ +# +# 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 transformers import AutoModelForSeq2SeqLM, AutoTokenizer +from bigdl.llm import optimize_model + +# you could tune the prompt based on your own model, +FLAN_T5_PROMPT_FORMAT = "<|User|>:{prompt}" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for flan-t5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="google/flan-t5-xxl", + help='The huggingface repo id for the flan-t5 model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="Translate to German: My name is Arthur", + 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 + model = AutoModelForSeq2SeqLM.from_pretrained(model_path, trust_remote_code=True) + + # With only one line to enable BigDL-LLM optimization on model + # "wo" module is not converted due to some issues of T5 model + # (https://github.com/huggingface/transformers/issues/20287), + # "lm_head" module is not converted to generate outputs with better quality + model = optimize_model(model, modules_to_not_convert=["wo", "lm_head"]) + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = FLAN_T5_PROMPT_FORMAT.format(prompt=args.prompt) + input_ids = tokenizer.encode(prompt, return_tensors="pt") + st = time.time() + 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) diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/README.md index 6f8fc22b..ec5db97f 100644 --- a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/README.md +++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/README.md @@ -23,6 +23,7 @@ You can use BigDL-LLM to run almost every Huggingface Transformer models with IN | Vicuna | [link](vicuna) | | Whisper | [link](whisper) | | Replit | [link](replit) | +| Flan-t5 | [link](flan-t5) | ## Verified Hardware Platforms diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/README.md new file mode 100644 index 00000000..a7d58fe5 --- /dev/null +++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/README.md @@ -0,0 +1,55 @@ +# Flan-t5 +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Flan-t5 models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) as a reference Flan-t5 model. + +## 0. Requirements +To run these examples with BigDL-LLM on Intel GPUs, 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 Flan-t5 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs. +### 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 + +# below command will install intel_extension_for_pytorch==2.0.110+xpu as default +# you can install specific ipex/torch version for your need +pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu +``` + +### 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 +``` + +```bash +python ./generate.py --prompt 'Translate to German: My name is Arthur' +``` + +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 Flan-t5 model (e.g. `google/flan-t5-xxl` to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/flan-t5-xxl'`. +- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Translate to German: My name is Arthur'`. +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`. + +#### Sample Output +#### [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +<|User|>:Translate to German: My name is Arthur +-------------------- Output -------------------- +Ich bin Arthur. +``` diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py new file mode 100644 index 00000000..8d6ec148 --- /dev/null +++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/flan-t5/generate.py @@ -0,0 +1,83 @@ +# +# 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 intel_extension_for_pytorch as ipex +import time +import argparse + +from bigdl.llm.transformers import AutoModelForSeq2SeqLM +from transformers import AutoTokenizer + +# you could tune the prompt based on your own model, +FLAN_T5_PROMPT_FORMAT = "<|User|>:{prompt}" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for flan-t5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="google/flan-t5-xxl", + help='The huggingface repo id for the flan-t5 model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="Translate to German: My name is Arthur", + 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. + # "wo" module is not converted due to some issues of T5 model + # (https://github.com/huggingface/transformers/issues/20287), + # "lm_head" module is not converted to generate outputs with better quality + model = AutoModelForSeq2SeqLM.from_pretrained(model_path, + load_in_4bit=True, + optimize_model=False, + trust_remote_code=True, + use_cache=True, + modules_to_not_convert=["wo", "lm_head"]) + model = model.to('xpu') + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = FLAN_T5_PROMPT_FORMAT.format(prompt=args.prompt) + input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu') + # ipex model needs a warmup, then inference time can be accurate + output = model.generate(input_ids, + max_new_tokens=args.n_predict) + + # start inference + 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) + torch.xpu.synchronize() + end = time.time() + output = output.cpu() + output_str = tokenizer.decode(output[0], skip_special_tokens=True) + output_str = output_str.split("")[0] + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + print(output_str) diff --git a/python/llm/example/GPU/PyTorch-Models/Model/README.md b/python/llm/example/GPU/PyTorch-Models/Model/README.md index ad74542c..41beaaee 100644 --- a/python/llm/example/GPU/PyTorch-Models/Model/README.md +++ b/python/llm/example/GPU/PyTorch-Models/Model/README.md @@ -13,6 +13,7 @@ You can use `optimize_model` API to accelerate general PyTorch models on Intel G | StarCoder | [link](starcoder) | | Dolly v1 | [link](dolly-v1) | | Dolly v2 | [link](dolly-v2) | +| Flan-t5 | [link](flan-t5) | ## Verified Hardware Platforms diff --git a/python/llm/example/GPU/PyTorch-Models/Model/flan-t5/README.md b/python/llm/example/GPU/PyTorch-Models/Model/flan-t5/README.md new file mode 100644 index 00000000..a7d58fe5 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/flan-t5/README.md @@ -0,0 +1,55 @@ +# Flan-t5 +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Flan-t5 models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) as a reference Flan-t5 model. + +## 0. Requirements +To run these examples with BigDL-LLM on Intel GPUs, 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 Flan-t5 model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations on Intel GPUs. +### 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 + +# below command will install intel_extension_for_pytorch==2.0.110+xpu as default +# you can install specific ipex/torch version for your need +pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu +``` + +### 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 +``` + +```bash +python ./generate.py --prompt 'Translate to German: My name is Arthur' +``` + +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 Flan-t5 model (e.g. `google/flan-t5-xxl` to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'google/flan-t5-xxl'`. +- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'Translate to German: My name is Arthur'`. +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`. + +#### Sample Output +#### [google/flan-t5-xxl](https://huggingface.co/google/flan-t5-xxl) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +<|User|>:Translate to German: My name is Arthur +-------------------- Output -------------------- +Ich bin Arthur. +``` diff --git a/python/llm/example/GPU/PyTorch-Models/Model/flan-t5/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/flan-t5/generate.py new file mode 100644 index 00000000..9dfbabc8 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/flan-t5/generate.py @@ -0,0 +1,78 @@ +# +# 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 intel_extension_for_pytorch as ipex +import time +import argparse + +from transformers import AutoModelForSeq2SeqLM, AutoTokenizer +from bigdl.llm import optimize_model + +# you could tune the prompt based on your own model, +FLAN_T5_PROMPT_FORMAT = "<|User|>:{prompt}" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for flan-t5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="google/flan-t5-xxl", + help='The huggingface repo id for the flan-t5 model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="Translate to German: My name is Arthur", + 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 + model = AutoModelForSeq2SeqLM.from_pretrained(model_path, + trust_remote_code=True, + torch_dtype='auto', + low_cpu_mem_usage=True) + + # With only one line to enable BigDL-LLM optimization on model + # "wo" module is not converted due to some issues of T5 model + # (https://github.com/huggingface/transformers/issues/20287), + # "lm_head" module is not converted to generate outputs with better quality + model = optimize_model(model, modules_to_not_convert=["wo", "lm_head"]) + + model = model.to('xpu') + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = FLAN_T5_PROMPT_FORMAT.format(prompt=args.prompt) + input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu') + # ipex model needs a warmup, then inference time can be accurate + output = model.generate(input_ids, + max_new_tokens=args.n_predict) + + # start inference + st = time.time() + output = model.generate(input_ids, + max_new_tokens=args.n_predict) + torch.xpu.synchronize() + end = time.time() + output = output.cpu() + 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)