LLM: add save/load example for hf-transformers (#10432)

JinBridger · web-flow · commit cc5806f4bc5c · 2024-03-22T13:57:47.000+08:00
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/.keep b/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/.keep
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/README.md b/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/README.md
@@ -0,0 +1,130 @@
+# Save/Load Low-Bit Models with BigDL-LLM Optimizations
+
+In this directory, you will find example on how you could save/load models with BigDL-LLM INT4 optimizations on Llama2 models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) and [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf) as reference Llama2 models.
+
+## 0. Requirements
+To run this example with BigDL-LLM, we have some recommended requirements for your machine, please refer to [here](../../README.md#system-support) for more information.
+
+## Example: Save/Load Model in Low-Bit Optimization
+In the example [generate.py](./generate.py), we show a basic use case of saving/loading model in low-bit optimizations to predict the next N tokens using `generate()` API. Also, saving and loading operations are platform-independent, so you could run it on different platforms.
+### 1. Install
+#### 1.1 Installation on Linux
+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==2.1.10+xpu as default
+pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+```
+
+#### 1.2 Installation on Windows
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9 libuv
+conda activate llm
+# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
+pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+```
+
+### 2. Configures OneAPI environment variables
+#### 2.1 Configurations for Linux
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+
+#### 2.2 Configurations for Windows
+```cmd
+call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
+```
+> Note: Please make sure you are using **CMD** (**Anaconda Prompt** if using conda) to run the command as PowerShell is not supported.
+
+
+### 3. Run
+#### 3.1 Configurations for Linux
+<details>
+
+<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
+
+```bash
+export USE_XETLA=OFF
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+```
+
+</details>
+
+<details>
+
+<summary>For Intel Data Center GPU Max Series</summary>
+
+```bash
+export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
+export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+export ENABLE_SDP_FUSION=1
+```
+> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
+</details>
+
+#### 3.2 Configurations for Windows
+<details>
+
+<summary>For Intel iGPU</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+set BIGDL_LLM_XMX_DISABLED=1
+```
+
+</details>
+
+<details>
+
+<summary>For Intel Arc™ A300-Series or Pro A60</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+<details>
+
+<summary>For other Intel dGPU Series</summary>
+
+There is no need to set further environment variables.
+
+</details>
+
+> Note: For the first time that each model runs on Intel iGPU/Intel Arc™ A300-Series or Pro A60, it may take several minutes to compile.
+
+### 4. Running examples
+
+If you want to save the optimized low-bit model, run:
+```
+python ./generate.py --save-path path/to/save/model
+```
+
+If you want to load the optimized low-bit model, run:
+```
+python ./generate.py --load-path path/to/load/model
+```
+
+In the example, several arguments can be passed to satisfy your requirements:
+
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Llama2 model to be downloaded, or the path to the ModelScope checkpoint folder. It is default to be `'meta-llama/Llama-2-7b-chat-hf'`.
+- `--save-path`: argument defining the path to save the low-bit model. Then you can load the low-bit directly.
+- `--load-path`: argument defining the path to load low-bit model.
+- `--prompt PROMPT`: argument defining the prompt to be inferred (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
+#### [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf)
+```log
+Inference time: xxxx s
+-------------------- Output --------------------
+### HUMAN:
+What is AI?
+
+### RESPONSE:
+
+AI is a term used to describe the development of computer systems that can perform tasks that typically require human intelligence, such as understanding natural language, recognizing images
+```
diff --git a/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/generate.py b/python/llm/example/GPU/HF-Transformers-AutoModels/Save-Load/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 time
+import argparse
+from bigdl.llm.transformers import AutoModelForCausalLM
+from transformers import LlamaTokenizer
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://huggingface.co/georgesung/llama2_7b_chat_uncensored#prompt-style
+LLAMA2_PROMPT_FORMAT = """### HUMAN:
+{prompt}
+
+### RESPONSE:
+"""
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Example of saving and loading the optimized model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="meta-llama/Llama-2-7b-chat-hf",
+                        help='The huggingface repo id for the Llama2 (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--save-path', type=str, default=None,
+                        help='The path to save the low-bit model.')
+    parser.add_argument('--load-path', type=str, default=None,
+                        help='The path to load the low-bit model.')
+    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_path = args.load_path
+    if load_path:
+        model = AutoModelForCausalLM.load_low_bit(load_path, trust_remote_code=True)
+        tokenizer = LlamaTokenizer.from_pretrained(load_path)
+    else:
+        model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                     load_in_4bit=True,
+                                                     trust_remote_code=True)
+        tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+    save_path = args.save_path
+    if save_path:
+        model.save_low_bit(save_path)
+        tokenizer.save_pretrained(save_path)
+        print(f"Model and tokenizer are saved to {save_path}")
+
+    # please save/load model before you run it on GPU
+    model = model.to('xpu')
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = LLAMA2_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)
+
+        st = time.time()
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict)
+        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, 'Output', '-'*20)
+        print(output_str)
