Add DeepSeek-MoE-16B-Chat (#10155)

* dsmoe-hf add

* add dsmoe pytorch

* update README

* modify comment

* remove GPU example

* update model name

* format code

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).***
 
 ---

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) |

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
+```
+
+

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)

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
+```
+

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)

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