From add3899311b3ef95c6ed0d59ccc89bc9709f69b4 Mon Sep 17 00:00:00 2001 From: Zhicun <59141989+ivy-lv11@users.noreply.github.com> Date: Tue, 20 Feb 2024 13:59:52 +0800 Subject: [PATCH] Add ziya CPU example (#10114) * ziya on CPU * add README for ziya * specify use_cache * add arc CPU * update prompt format * update link * add comments to emphasize use_cache * update pip cmd --- README.md | 1 + python/llm/README.md | 2 +- .../Model/ziya/README.md | 88 +++++++++++++++++++ .../Model/ziya/generate.py | 77 ++++++++++++++++ .../CPU/PyTorch-Models/Model/ziya/README.md | 78 ++++++++++++++++ .../CPU/PyTorch-Models/Model/ziya/generate.py | 78 ++++++++++++++++ 6 files changed, 323 insertions(+), 1 deletion(-) create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/README.md create mode 100644 python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/generate.py create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/ziya/README.md create mode 100644 python/llm/example/CPU/PyTorch-Models/Model/ziya/generate.py diff --git a/README.md b/README.md index f1447419ace..80ccad311f9 100644 --- a/README.md +++ b/README.md @@ -185,6 +185,7 @@ Over 40 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | RWKV5 | | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/rwkv5) | | Bark | [link](python/llm/example/CPU/PyTorch-Models/Model/bark) | [link](python/llm/example/GPU/PyTorch-Models/Model/bark) | | SpeechT5 | | [link](python/llm/example/GPU/PyTorch-Models/Model/speech-t5) | +| Ziya-Coding-34B-v1.0 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya) | | ***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 be38d2a0892..6e765d390c8 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -81,7 +81,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | RWKV5 | | [link](example/GPU/HF-Transformers-AutoModels/Model/rwkv5) | | Bark | [link](example/CPU/PyTorch-Models/Model/bark) | [link](example/GPU/PyTorch-Models/Model/bark) | | SpeechT5 | | [link](example/GPU/PyTorch-Models/Model/speech-t5) | - +| Ziya-Coding-34B-v1.0 | [link](example/CPU/HF-Transformers-AutoModels/Model/ziya) | | ### Working with `bigdl-llm` diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/README.md new file mode 100644 index 00000000000..e1f79120703 --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/README.md @@ -0,0 +1,88 @@ +# Ziya + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on Ziya models. For illustration purposes, we utilize the [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) as a reference Ziya 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 Ziya 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 Ziya 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 Ziya 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 'def quick_sort(arr):\n' +``` +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-LLM env variables +source bigdl-llm-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 'def quick_sort(arr):\n' +``` +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 Ziya model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'IDEA-CCNL/Ziya-Coding-34B-v1.0'`. +- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `def quick_sort(arr):\n`. +- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `128`. + +#### 2.4 Sample Output +#### [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +: +def quick_sort(arr):\n +: + +-------------------- Output -------------------- + : +def quick_sort(arr):\n +: +def partition(arr, low, high): + + i = (low-1) + pivot = arr[high] + for j in range(low, high): + if arr[j] <= pivot: + arr[i], arr[j] = arr[j], arr[i] + i = i+1 + arr[i], arr[high] = arr[high], arr[i] + return i + +def quick_sort(arr, low, high): + if low < high: + pi = partition(arr, low, +``` diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/generate.py new file mode 100644 index 00000000000..ab5708e83c4 --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya/generate.py @@ -0,0 +1,77 @@ +# +# 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 transformers import AutoTokenizer + +# you could tune the prompt based on your own model, +# here the prompt tuning refers to https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0 +ZIYA_PROMPT_FORMAT = ": \n{prompt}\n: \n" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Ziya model') + parser.add_argument('--repo-id-or-model-path', type=str, default="IDEA-CCNL/Ziya-Coding-34B-v1.0", + help='The huggingface repo id for the Ziya model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="def quick_sort(arr):\n", + help='Prompt to infer') + parser.add_argument('--n-predict', type=int, default=128, + help='Max tokens to predict') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + + + from bigdl.llm.transformers import AutoModelForCausalLM + # enabling `use_cache=True` allows the model to utilize the previous + # key/values attentions to speed up decoding; + # to obtain optimal performance with BigDL-LLM INT4 optimizations, + # it is important to set use_cache=True for Ziya models + model = AutoModelForCausalLM.from_pretrained(model_path, + load_in_4bit=True, + trust_remote_code=True, + use_cache=True) + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = ZIYA_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, + do_sample = True, + top_p = 0.85, + temperature = 0.8, + repetition_penalty = 0.95, + eos_token_id = tokenizer.eos_token_id, + pad_token_id = tokenizer.pad_token_id, + ) + end = time.time() + output_str = tokenizer.batch_decode(output)[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/ziya/README.md b/python/llm/example/CPU/PyTorch-Models/Model/ziya/README.md new file mode 100644 index 00000000000..2b6dadcd5db --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/ziya/README.md @@ -0,0 +1,78 @@ +# Ziya +In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate Ziya models. For illustration purposes, we utilize the [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) as a reference Ziya 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 Ziya model to predict the next N tokens using `generate()` API, with BigDL-LLM 'optimize_model' API. +### 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 Ziya to conduct generation +``` + +### 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 'def quick_sort(arr):\n' +``` +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-LLM env variables +source bigdl-llm-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 +``` +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 Ziya model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'IDEA-CCNL/Ziya-Coding-34B-v1.0'`. +- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be `def quick_sort(arr):\n`. +- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `128`. + +#### 2.4 Sample Output +#### [IDEA-CCNL/Ziya-Coding-34B-v1.0](https://huggingface.co/IDEA-CCNL/Ziya-Coding-34B-v1.0) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +: +def quick_sort(arr):\n +: + +-------------------- Output -------------------- + : +def quick_sort(arr):\n +: +def partition(arr, low, high): + + i = (low-1) + pivot = arr[high] + for j in range(low, high): + if arr[j] <= pivot: + arr[i], arr[j] = arr[j], arr[i] + i = i+1 + arr[i], arr[high] = arr[high], arr[i] + return i + +def quick_sort(arr, low, high): + if low < high: + pi = partition(arr, low, +``` diff --git a/python/llm/example/CPU/PyTorch-Models/Model/ziya/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/ziya/generate.py new file mode 100644 index 00000000000..1604695125b --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/ziya/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 time +import argparse +import numpy as np + +from transformers import AutoTokenizer + + +ZIYA_PROMPT_FORMAT = ": \n{prompt}\n: \n" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Ziya model') + parser.add_argument('--repo-id-or-model-path', type=str, default="IDEA-CCNL/Ziya-Coding-34B-v1.0", + help='The huggingface repo id for the Ziya model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="def quick_sort(arr):\n", + help='Prompt to infer') + parser.add_argument('--n-predict', type=int, default=128, + help='Max tokens to predict') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + + + from transformers import AutoModelForCausalLM + from bigdl.llm import optimize_model + # enabling `use_cache=True` allows the model to utilize the previous + # key/values attentions to speed up decoding; + # to obtain optimal performance with BigDL-LLM `optimization_model` API optimizations, + # it is important to set use_cache=True for Ziya models + model = AutoModelForCausalLM.from_pretrained(model_path, + trust_remote_code=True, + use_cache=True) + model = optimize_model(model) + + # Load tokenizer + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + prompt = ZIYA_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, + do_sample = True, + top_p = 0.85, + temperature = 0.8, + repetition_penalty = 0.95, + eos_token_id = tokenizer.eos_token_id, + pad_token_id = tokenizer.pad_token_id, + ) + end = time.time() + output_str = tokenizer.batch_decode(output)[0] + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + print(output_str) + \ No newline at end of file