diff --git a/README.md b/README.md index 4f4765fb..f48d36cf 100644 --- a/README.md +++ b/README.md @@ -189,12 +189,14 @@ 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) | +| DeepSeek-MoE | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe) | | | Ziya-Coding-34B-v1.0 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/ziya) | | | Phi-2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/phi-2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/phi-2) | | Yuan2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/yuan2) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/yuan2) | | Gemma | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/gemma) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/gemma) | | DeciLM-7B | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/deciLM-7b) | [link](python/llm/example/GPU/HF-Transformers-AutoModels/Model/deciLM-7b) | + ***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 881c09f0..9d4d87b5 100644 --- a/python/llm/README.md +++ b/python/llm/README.md @@ -81,6 +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) | +| DeepSeek-MoE | [link](example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe) | | | Ziya-Coding-34B-v1.0 | [link](example/CPU/HF-Transformers-AutoModels/Model/ziya) | | | Phi-2 | [link](example/CPU/HF-Transformers-AutoModels/Model/phi-2) | [link](example/GPU/HF-Transformers-AutoModels/Model/phi-2) | | Yuan2 | [link](example/CPU/HF-Transformers-AutoModels/Model/yuan2) | [link](example/GPU/HF-Transformers-AutoModels/Model/yuan2) | diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe/README.md b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe/README.md new file mode 100644 index 00000000..a396693f --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe/README.md @@ -0,0 +1,73 @@ +# DeepSeek-MoE + +In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on DeepSeek-MoE models. For illustration purposes, we utilize the [deepseek-ai/deepseek-moe-16b-chat](https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat) as a reference DeepSeek-MoE 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 DeepSeek-MoE 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 DeepSeek-MoE 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 DeepSeek-MoE 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-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 '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 DeepSeek-MoE model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'deepseek-ai/DeepSeek-MoE-16b-chat'`. +- `--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 +#### [deepseek-ai/DeepSeek-MoE-16b-chat](https://huggingface.co/deepseek-ai/DeepSeek-MoE-16b-chat) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +<｜begin▁of▁sentence｜>[INST] What is AI? [/INST] +-------------------- Output -------------------- +<｜begin▁of▁sentence｜>[INST] What is AI? [/INST] + +Artificial Intelligence (AI) is a field of computer science that aims to create intelligent machines that can perform tasks that typically require human intelligence, such as learning +``` + + diff --git a/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe/generate.py b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe/generate.py new file mode 100644 index 00000000..87cca75f --- /dev/null +++ b/python/llm/example/CPU/HF-Transformers-AutoModels/Model/deepseek-moe/generate.py @@ -0,0 +1,62 @@ +# +# 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, GenerationConfig + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for DeepSeek-MoE model') + parser.add_argument('--repo-id-or-model-path', type=str, default="/mnt/disk1/models/deepseek-moe-16b-chat", + help='The huggingface repo id for the CodeShell 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 + from bigdl.llm.transformers import AutoModelForCausalLM + model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, load_in_4bit=True).eval() + model.generation_config = GenerationConfig.from_pretrained(model_path) + model.generation_config.pad_token_id = model.generation_config.eos_token_id + + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + messages = [ + {"role": "user", "content": args.prompt} + ] + prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) + input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") + st = time.time() + outputs = model.generate(input_tensor.to(model.device), max_new_tokens=args.n_predict) + end = time.time() + result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True) + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + print(result) diff --git a/python/llm/example/CPU/PyTorch-Models/Model/deepseek-moe/README.md b/python/llm/example/CPU/PyTorch-Models/Model/deepseek-moe/README.md new file mode 100644 index 00000000..adfc8006 --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/deepseek-moe/README.md @@ -0,0 +1,63 @@ +# DeepSeek-MoE +In this directory, you will find examples on how you could use BigDL-LLM `optimize_model` API to accelerate DeepSeek-MoE models. For illustration purposes, we utilize the [deepseek-ai/DeepSeek-MoE-16b-chat](https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat) as a reference DeepSeek-MoE 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 deepseek-moe 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-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 '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 DeepSeek-MoE model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'deepseek-ai/deepseek-moe-16b-chat'`. +- `--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 +#### [deepseek-ai/deepseek-moe-16b-chat](https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat) +```log +Inference time: xxxx s +-------------------- Prompt -------------------- +<｜begin▁of▁sentence｜>[INST] What is AI? [/INST] +-------------------- Output -------------------- +<｜begin▁of▁sentence｜>[INST] What is AI? [/INST] + +Artificial Intelligence (AI) is a field of computer science that aims to create intelligent machines that can perform tasks that typically require human intelligence, such as learning +``` + diff --git a/python/llm/example/CPU/PyTorch-Models/Model/deepseek-moe/generate.py b/python/llm/example/CPU/PyTorch-Models/Model/deepseek-moe/generate.py new file mode 100644 index 00000000..d0ca949a --- /dev/null +++ b/python/llm/example/CPU/PyTorch-Models/Model/deepseek-moe/generate.py @@ -0,0 +1,62 @@ +# +# 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, GenerationConfig + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for DeepSeek-MoE model') + parser.add_argument('--repo-id-or-model-path', type=str, default="/mnt/disk1/models/deepseek-moe-16b-chat", + help='The huggingface repo id for the CodeShell 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 + + + from transformers import AutoModelForCausalLM + from bigdl.llm import optimize_model + model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype = torch.bfloat16, device_map = "auto", attn_implementation="eager") + model.generation_config = GenerationConfig.from_pretrained(model_path) + model.generation_config.pad_token_id = model.generation_config.eos_token_id + model = optimize_model(model) + tokenizer = AutoTokenizer.from_pretrained(model_path, + trust_remote_code=True) + + # Generate predicted tokens + with torch.inference_mode(): + messages = [ + {"role": "user", "content": args.prompt} + ] + prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) + input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") + st = time.time() + outputs = model.generate(input_tensor.to(model.device), max_new_tokens=args.n_predict) + end = time.time() + result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True) + print(f'Inference time: {end-st} s') + print('-'*20, 'Prompt', '-'*20) + print(prompt) + print('-'*20, 'Output', '-'*20) + print(result) diff --git a/python/llm/src/bigdl/llm/transformers/convert.py b/python/llm/src/bigdl/llm/transformers/convert.py index bcf08a93..0717523d 100644 --- a/python/llm/src/bigdl/llm/transformers/convert.py +++ b/python/llm/src/bigdl/llm/transformers/convert.py @@ -861,6 +861,7 @@ def _optimize_post(model, lightweight_bmm=False): # transformers version >= 4.36.0 from bigdl.llm.transformers.models.falcon import \ falcon_attention_forward_4_36 + convert_forward(model, module.FalconAttention, falcon_attention_forward_4_36