From 4a19f50d16c9e5b084df109e74f43da7a8a77264 Mon Sep 17 00:00:00 2001 From: Yining Wang Date: Tue, 24 Oct 2023 15:08:04 +0800 Subject: [PATCH] phi-1_5 CPU and GPU examples (#9173) * eee * add examples on CPU and GPU * fix * fix * optimize model examples * have updated * Warmup and configs added * Update two tables --- README.md | 1 + python/llm/README.md | 26 ++++-- .../Model/phi-1_5/README.md | 74 +++++++++++++++++ .../Model/phi-1_5/generate.py | 72 ++++++++++++++++ .../PyTorch-Models/Model/phi-1_5/README.md | 60 ++++++++++++++ .../PyTorch-Models/Model/phi-1_5/generate.py | 61 ++++++++++++++ .../Model/phi-1_5/README.md | 56 +++++++++++++ .../Model/phi-1_5/generate.py | 82 +++++++++++++++++++ .../PyTorch-Models/Model/phi-1_5/README.md | 53 ++++++++++++ .../PyTorch-Models/Model/phi-1_5/generate.py | 72 ++++++++++++++++ 10 files changed, 549 insertions(+), 8 deletions(-) create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/README.md create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/generate.py create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md create mode 100644 python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/README.md create mode 100644 python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/generate.py diff --git a/README.md b/README.md index 33aa7e00..1b384ebe 100644 --- a/README.md +++ b/README.md @@ -151,6 +151,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | Aquila | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/aquila) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/aquila) | | 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) | ***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 92ec5c11..7e1c2233 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -58,20 +58,30 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | Aquila | [link](example/CPU/HF-Transformers-AutoModels/Model/aquila) | [link](example/GPU/HF-Transformers-AutoModels/Model/aquila) | | 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) | ### Working with `bigdl-llm`
Table of Contents -- [Install](#install) -- [Run Model](#run-model) - - [Hugging Face `transformers` API](#1-hugging-face-transformers-api) - - [Native INT4 Model](#2-native-int4-model) - - [LangChain API](#l3-angchain-api) - - [CLI Tool](#4-cli-tool) -- [`bigdl-llm` API Doc](#bigdl-llm-api-doc) -- [`bigdl-llm` Dependency](#bigdl-llm-dependency) +- [BigDL-LLM](#bigdl-llm) + - [Demos](#demos) + - [Verified models](#verified-models) + - [Working with `bigdl-llm`](#working-with-bigdl-llm) + - [Install](#install) + - [CPU](#cpu) + - [GPU](#gpu) + - [Run Model](#run-model) + - [1. Hugging Face `transformers` API](#1-hugging-face-transformers-api) + - [CPU INT4](#cpu-int4) + - [GPU INT4](#gpu-int4) + - [More Low-Bit Support](#more-low-bit-support) + - [2. Native INT4 model](#2-native-int4-model) + - [3. LangChain API](#3-langchain-api) + - [4. CLI Tool](#4-cli-tool) + - [`bigdl-llm` API Doc](#bigdl-llm-api-doc) + - [`bigdl-llm` Dependency](#bigdl-llm-dependency)
diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md new file mode 100644 index 00000000..013704c8 --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md @@ -0,0 +1,74 @@ +# phi-1_5 + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on phi-1_5 models. For illustration purposes, we utilize the [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) as a reference phi-1_5 model. + +> **Note**: If you want to download the Hugging Face *Transformers* model, please refer to [here](https://huggingface.co/docs/hub/models-downloading#using-git). +> +> BigDL-LLM optimizes the *Transformers* model in INT4 precision at runtime, and thus no explicit conversion is needed. + +## 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 phi-1_5 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 einops # additional package required for phi-1_5 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 phi-1_5 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 'What is AI?' +``` +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 'What is AI？' +``` +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`: str, argument defining the huggingface repo id for the phi-1_5 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-1_5'`. +- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `What is AI?`. +- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`. + +#### 2.4 Sample Output +#### [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +Question: What is AI? + + Answer: +-------------------- Output -------------------- +Question: What is AI? + + Answer: AI stands for Artificial Intelligence, which refers to the development of computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, + +``` diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py new file mode 100644 index 00000000..e934d99b --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py @@ -0,0 +1,72 @@ +# +# 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 AutoModel,AutoModelForCausalLM +from transformers import AutoTokenizer, GenerationConfig + +# you could tune the prompt based on your own model, +# here the prompt tuning refers to # TODO: https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py +PHI1_5_PROMPT_FORMAT = " Question:{prompt}\n\n Answer:" +generation_config = GenerationConfig(use_cache = True) + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-1_5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-1_5", + help='The huggingface repo id for the phi-1_5 model 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 = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = PHI1_5_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 + + # Note that phi-1_5 uses GenerationConfig to enable 'use_cache' + output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config) + + 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/CPU/PyTorch-Models/Model/phi-1_5/README.md b/python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/README.md new file mode 100644 index 00000000..f925d0d6 --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/README.md @@ -0,0 +1,60 @@ +# phi-1_5 +In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate phi-1_5 models. For illustration purposes, we utilize the [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) as a reference phi-1_5 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 a phi-1_5 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 einops +``` + +### 2. Run +After setting up the Python environment, you could run the example by following steps. + +#### 2.1 Client +On client Windows machines, it is recommended to run directly with full utilization of all cores: +```powershell +python ./generate.py --prompt 'What is AI?' +``` +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 'What is AI?' +``` +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`: str, argument defining the huggingface repo id for the phi-1_5 model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-1_5'`. +- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `'What is AI?'`. +- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `32`. + +#### 2.4 Sample Output +#### [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) +```log +Inference time: xxxx s +-------------------- Output -------------------- +Question: What is AI? + +Answer: AI stands for Artificial Intelligence, which refers to the development of computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, +``` diff --git a/python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/generate.py new file mode 100644 index 00000000..f819a47f --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/phi-1_5/generate.py @@ -0,0 +1,61 @@ +# +# 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 AutoModel, AutoTokenizer, AutoModelForCausalLM, GenerationConfig +from bigdl.llm import optimize_model + +# you could tune the prompt based on your own model, +# here the prompt tuning refers to https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py +PHI_1_5_V1_PROMPT_FORMAT = "Question: {prompt}\n\n Answer:" +generation_config = GenerationConfig(use_cache = True) + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-1_5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-1_5", + help='The huggingface repo id for the phi-1_5 model 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 + model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True) + + # With only one line to enable BigDL-LLM optimization on model + model = optimize_model(model) + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = PHI_1_5_V1_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, generation_config = generation_config) + end = time.time() + output_str = tokenizer.decode(output[0], skip_special_tokens=True) + print(f'Inference time: {end-st} s') + print('-'*20, 'Output', '-'*20) + print(output_str) diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md new file mode 100644 index 00000000..09d07df7 --- /dev/null +++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/README.md @@ -0,0 +1,56 @@ +# phi-1_5 +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on phi-1_5 models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) as a reference phi-1_5 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 phi-1_5 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 environment: +```bash +conda create -n llm 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 +pip install einops # additional package required for phi-1_5 to conduct generation +``` + +### 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 +``` + +``` +python ./generate.py --prompt 'What is AI?' +``` + +Arguments info: +- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the phi-1_5 model (e.g. `microsoft/phi-1_5`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-1_5'`. +- `--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`. + +#### Sample Output +#### [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +Question: What is AI? + + Answer: +-------------------- Output -------------------- +Question: What is AI? + + Answer: AI stands for Artificial Intelligence, which refers to the development of computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, +``` diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py new file mode 100644 index 00000000..2108654d --- /dev/null +++ b/python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-1_5/generate.py @@ -0,0 +1,82 @@ +# +# 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 +import numpy as np + +from bigdl.llm.transformers import AutoModel,AutoModelForCausalLM +from transformers import AutoTokenizer, GenerationConfig + +# you could tune the prompt based on your own model, +# here the prompt tuning refers to # TODO: https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py +PHI1_5_PROMPT_FORMAT = " Question:{prompt}\n\n Answer:" +generation_config = GenerationConfig(use_cache = True) + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-1_5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-1_5", + help='The huggingface repo id for the phi-1_5 model 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) + + model = model.to('xpu') + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = PHI1_5_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, + generation_config = generation_config) + # 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 + + # Note that phi-1_5 uses GenerationConfig to enable 'use_cache' + output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config) + 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) diff --git a/python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/README.md b/python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/README.md new file mode 100644 index 00000000..0c1d0a76 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/README.md @@ -0,0 +1,53 @@ +# phi-1_5 +In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate phi-1_5 models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) as a reference phi-1_5 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 a phi-1_5 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 + +pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu +pip install einops # additional package required for phi-1_5 to conduct generation +``` + +### 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 +``` + +``` +python ./generate.py --prompt 'What is AI?' +``` + +Arguments info: +- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the phi-1_5 model (e.g. `microsoft/phi-1_5`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'microsoft/phi-1_5'`. +- `--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`. + +#### Sample Output +#### [microsoft/phi-1_5](https://huggingface.co/microsoft/phi-1_5) + +```log +Inference time: xxxx s +-------------------- Output -------------------- +Question: What is AI? + +Answer: AI stands for Artificial Intelligence, which refers to the development of computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, +``` \ No newline at end of file diff --git a/python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/generate.py new file mode 100644 index 00000000..f5f2bae0 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/phi-1_5/generate.py @@ -0,0 +1,72 @@ +# +# 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 AutoModel, AutoTokenizer, AutoModelForCausalLM, GenerationConfig +from bigdl.llm import optimize_model + +# you could tune the prompt based on your own model, +# here the prompt tuning refers to https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_mixformer_sequential.py +PHI_1_5_V1_PROMPT_FORMAT = "Question: {prompt}\n\n Answer:" +generation_config = GenerationConfig(use_cache = True) + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-1_5 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-1_5", + help='The huggingface repo id for the phi-1_5 model 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 + model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True) + + # With only one line to enable BigDL-LLM optimization on model + model = optimize_model(model) + model = model.to('xpu') + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = PHI_1_5_V1_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, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config) + # start inference + st = time.time() + # Note that phi-1_5 uses GenerationConfig to enable 'use_cache' + output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config) + 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)