diff --git a/README.md b/README.md index 5aa58129..e7e73a56 100644 --- a/README.md +++ b/README.md @@ -181,6 +181,7 @@ Over 40 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | Fuyu | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/fuyu) | | | Distil-Whisper | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/distil-whisper) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/distil-whisper) | | Yi | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/yi) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/yi) | +| RWKV | [link](python/llm/example/CPU/PyTorch-Models/Model/rwkv) | [link](python/llm/example/GPU/PyTorch-Models/Model/rwkv) | | BlueLM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/bluelm) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/bluelm) | | Mamba | [link](python/llm/example/CPU/PyTorch-Models/Model/mamba) | [link](python/llm/example/GPU/PyTorch-Models/Model/mamba) | | SOLAR | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/solar) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/solar) | diff --git a/python/llm/README.md b/python/llm/README.md index 4ddced57..a8700dd3 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -73,6 +73,7 @@ Over 20 models have been optimized/verified on `bigdl-llm`, including *LLaMA/LLa | Fuyu | [link](example/CPU/HF-Transformers-AutoModels/Model/fuyu) | | | Distil-Whisper | [link](example/CPU/HF-Transformers-AutoModels/Model/distil-whisper) | [link](example/GPU/HF-Transformers-AutoModels/Model/distil-whisper) | | Yi | [link](example/CPU/HF-Transformers-AutoModels/Model/yi) | [link](example/GPU/HF-Transformers-AutoModels/Model/yi) | +| RWKV | [link](example/CPU/PyTorch-Models/Model/rwkv) | [link](example/GPU/PyTorch-Models/Model/rwkv) | | BlueLM | [link](example/CPU/HF-Transformers-AutoModels/Model/bluelm) | [link](example/GPU/HF-Transformers-AutoModels/Model/bluelm) | | Mamba | [link](example/CPU/PyTorch-Models/Model/mamba) | [link](example/GPU/PyTorch-Models/Model/mamba) | | SOLAR | [link](example/CPU/HF-Transformers-AutoModels/Model/solar) | [link](example/GPU/HF-Transformers-AutoModels/Model/solar) | diff --git a/python/llm/example/CPU/PyTorch-Models/Model/rwkv/README.md b/python/llm/example/CPU/PyTorch-Models/Model/rwkv/README.md new file mode 100644 index 00000000..1cf6c8b1 --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/rwkv/README.md @@ -0,0 +1,75 @@ +# RWKV + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on RWKV models. For illustration purposes, we utilize the [RWKV/rwkv-4-world-7b](https://huggingface.co/RWKV/rwkv-4-world-7b) as a reference RWKV 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 RWKV 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 RWKV 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 "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley" +``` +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 "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley" +``` +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 RWKV model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'RWKV/rwkv-4-world-7b'`. +- `--prompt`: str, argument defining the prompt to be inferred (with integrated prompt format for chat). It is default to be "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley". +- `--n-predict`: int, argument defining the max number of tokens to predict. It is default to be `40`. + +#### 2.4 Sample Output +#### [RWKV/rwkv-4-world-7b](https://huggingface.co/RWKV/rwkv-4-world-7b) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +Question: +In a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese. + +Answer: +-------------------- Output -------------------- +Question: +In a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese. + +Answer: 科学家在一个不为人知的谷地发现一群能说中文的龙。科学家惊讶地发现这些龙是中国的 + +``` diff --git a/python/llm/example/CPU/PyTorch-Models/Model/rwkv/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/rwkv/generate.py new file mode 100644 index 00000000..162d6b2e --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/rwkv/generate.py @@ -0,0 +1,71 @@ +# +# Copyright 2016 The BigDL Authors. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or ag8reed 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 +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/RWKV/rwkv-4-world-7b +RWKV_PROMPT_FORMAT = "Question: {prompt}\n\nAnswer:" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for RWKV model') + parser.add_argument('--repo-id-or-model-path', type=str, default="RWKV/rwkv-4-world-7b", + help='The huggingface repo id for the RWKV model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese.", + help='Prompt to infer') + parser.add_argument('--n-predict', type=int, default=40, + help='Max tokens to predict') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + + # First load the model in fp16 dtype + model = AutoModelForCausalLM.from_pretrained(model_path, + trust_remote_code=True, + low_cpu_mem_usage=True, + torch_dtype=torch.half) + + # Call the `_rescale_layers` method, prepare to convert to int4 + model.rwkv._rescale_layers() + + # 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 = RWKV_PROMPT_FORMAT.format(prompt = args.prompt) + + inputs = tokenizer(prompt, return_tensors="pt") + st = time.time() + output = model.generate(inputs["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/PyTorch-Models/Model/replit/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/replit/generate.py index 57f2c9f6..52ef1a27 100644 --- a/python/llm/example/GPU/PyTorch-Models/Model/replit/generate.py +++ b/python/llm/example/GPU/PyTorch-Models/Model/replit/generate.py @@ -69,6 +69,7 @@ if __name__ == '__main__': 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, 'Output', '-'*20) print(output_str) diff --git a/python/llm/example/GPU/PyTorch-Models/Model/rwkv/README.md b/python/llm/example/GPU/PyTorch-Models/Model/rwkv/README.md new file mode 100644 index 00000000..7ce117ee --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/rwkv/README.md @@ -0,0 +1,59 @@ +# RWKV + +In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate RWKV models on [Intel GPUs](../README.md). For illustration purposes, we utilize the [RWKV/rwkv-4-world-7b](https://huggingface.co/RWKV/rwkv-4-world-7b) as a reference RWKV 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 RWKV 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 +``` + +### 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 "你叫什么名字？" +``` + +Arguments info: +- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the RWKV model (e.g. `RWKV/rwkv-4-world-7b`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'RWKV/rwkv-4-world-7b'`. +- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `"你叫什么名字？"`. +- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `40`. + +#### Sample Output +#### [RWKV/rwkv-4-world-7b](https://huggingface.co/RWKV/rwkv-4-world-7b) + +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +Question: 你叫什么名字？ + +Answer: +-------------------- Output -------------------- +Question: 你叫什么名字？ + +Answer: 我是一个大型语言模型，没有具体的姓名。我是由OpenAI团队创建的，目的是为了提供自然 + + +``` diff --git a/python/llm/example/GPU/PyTorch-Models/Model/rwkv/generate.py b/python/llm/example/GPU/PyTorch-Models/Model/rwkv/generate.py new file mode 100644 index 00000000..e9ab7407 --- /dev/null +++ b/python/llm/example/GPU/PyTorch-Models/Model/rwkv/generate.py @@ -0,0 +1,80 @@ +# +# 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 +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/RWKV/rwkv-4-world-7b + +RWKV_PROMPT_FORMAT = "Question: {prompt}\n\nAnswer:" + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for RWKV model') + parser.add_argument('--repo-id-or-model-path', type=str, default="RWKV/rwkv-4-world-7b", + help='The huggingface repo id for the RWKV model to be downloaded' + ', or the path to the huggingface checkpoint folder') + parser.add_argument('--prompt', type=str, default="你叫什么名字？", + help='Prompt to infer') + parser.add_argument('--n-predict', type=int, default=40, + help='Max tokens to predict') + + args = parser.parse_args() + model_path = args.repo_id_or_model_path + + # First load the model in fp16 dtype + model = AutoModelForCausalLM.from_pretrained(model_path, + trust_remote_code=True, + low_cpu_mem_usage=True, + torch_dtype=torch.half) + + # Call the `_rescale_layers` method, prepare to convert to int4 + model.rwkv._rescale_layers() + + # 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 = RWKV_PROMPT_FORMAT.format(prompt = args.prompt) + inputs = tokenizer(prompt, return_tensors="pt").to('xpu') + + # ipex model needs a warmup, then inference time can be accurate + output = model.generate(inputs["input_ids"], + max_new_tokens=args.n_predict) + + # start inference + st = time.time() + output = model.generate(inputs["input_ids"], + max_new_tokens=args.n_predict) + torch.xpu.synchronize() + output = output.cpu() + 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)