Quick start: Install bigdl-llm on windows gpu

docs/readthedocs/source/doc/LLM/QuickStart/Install BigDL-LLM on Windows for Intel GPU.md

@@ -0,0 +1,157 @@
+# Install BigDL-LLM on Windows for Intel GPU
+
+## iGPU
+
+### Install GPU driver
+
+1. Step 1: Install Visual Studio 2022 Community Edition from [here](https://visualstudio.microsoft.com/downloads/). 
+
+   <img src="./figs/fig1.png"  style="zoom:20%;" />
+
+> Note select `Desktop development with C++` during installation. 
+>
+> <img src="./figs/fig2.png" alt="image-20240221102252560" style="zoom:40%;" />
+>
+> The installation could be slow and cost 15 minutes. Need at least 7GB. 
+>
+> If you do not select this workload during installation, go to Tools > Get Tools and Features... to change workload following [this page](https://learn.microsoft.com/en-us/cpp/build/vscpp-step-0-installation?view=msvc-170#step-4---choose-workloads). 
+
+2. Step 2: Install latest GPU driver from [here](https://www.intel.com/content/www/us/en/download/785597/intel-arc-iris-xe-graphics-windows.html). Note the process could be slow and It takes 10 minutes to download and install. Reboot is also needed.  After rebooting, we can check GPU status from GUI. 
+
+<img src="./figs/fig3.png" alt="image-20240221102217795" style="zoom:20%;" />
+
+<img src="./figs/fig4.png" alt="image-20240221105834031" style="zoom:20%;" />
+
+### Install conda
+
+We recommend using miniconda to create environment. Please refer to the [page](https://docs.anaconda.com/free/miniconda/) to install miniconda. 
+
+* Choose windows miniconda installer. Download and install. It takes a few minutes. 
+
+  <img src="./figs/fig5.png" alt="image-20240221110402278" style="zoom:20%;" />
+
+* After installation, open `Anaconda prompt` and create an environment by `conda create -n llm python=3.9 libuv` . 
+
+  > Note: if you encounter CondaHTTPError problem and fail to create the environment, please check the internet connection and proxy setting. You can define your proxy setting by `conda config --set proxy_servers.http your_http_proxy_IP:port` and `conda config --set proxy_servers.https your_https_proxy_IP:port`
+  >
+  > <img src="./figs/fig6.png" alt="image-20240221122852777" style="zoom:20%;" />
+
+### Install oneAPI 
+
+ Install oneAPI Base Toolkit with the help of pip. After ensuring  `conda` is ready, we can use `pip ` to install oneAPI Base Toolkit. 
+
+```bash
+pip install dpcpp-cpp-rt==2024.0.2 mkl-dpcpp==2024.0.0 onednn==2024.0.0
+```
+
+> If you encounter HTTP Timeout error, also check your internet and proxy setting in `pip.ini` file which is under "C:\Users\YourName\AppData\Roaming\pip"  folder. 
+
+When you successfully install oneAPI from pip, you will see similar thing from the Anaconda prompt command line. <img src="./figs/fig7.png" alt="image-20240221130508668" style="zoom:20%;" />
+
+### Install bigdl-llm
+
+1. Step 1: Run the commands below in Anaconda prompt. 
+
+```bash
+conda create -n llm python=3.9 libuv # Already done in "Install conda" section
+conda activate llm
+pip install dpcpp-cpp-rt==2024.0.2 mkl-dpcpp==2024.0.0 onednn==2024.0.0 # Already done in "Install oneAPI" section
+pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu
+```
+
+
+2. Step 2: Now we can test whether all the components have been installed correctly. If we can import all the packages correctly following the python file below, then the installation is correct. 
+   ```python
+   import torch
+   import time
+   import argparse
+   import numpy as np
+
+   from bigdl.llm.transformers import AutoModel,AutoModelForCausalLM
+   from transformers import AutoTokenizer, GenerationConfig
+   ```
+
+   Then we use phi-1.5 as an example to show how to run the model with bigdl-llm on windows. 
+   ```python
+   import torch
+   import time
+   import argparse
+   import numpy as np
+
+   from bigdl.llm.transformers import AutoModel,AutoModelForCausalLM
+   from transformers import AutoTokenizer, GenerationConfig
+
+   PHI1_5_PROMPT_FORMAT = " Question:{prompt}\n\n Answer:"
+   generation_config = GenerationConfig(use_cache = True)
+
+   if __name__ == '__main__':
+       parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for phi-1_5 model')
+       parser.add_argument('--repo-id-or-model-path', type=str, default="microsoft/phi-1_5",
+                           help='The huggingface repo id for the phi-1_5 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
+       # When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
+       # This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
+       model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                    load_in_4bit=True,
+                                                    cpu_embedding=True,
+                                                    trust_remote_code=True)
+
+       model = model.to('xpu')
+
+       # Load tokenizer
+       tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                                 trust_remote_code=True)
+
+       # Generate predicted tokens
+       with torch.inference_mode():
+           prompt = PHI1_5_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,
+                                   generation_config = generation_config)
+           # start inference
+           st = time.time()
+           # if your selected model is capable of utilizing previous key/value attentions
+           # to enhance decoding speed, but has `"use_cache": false` in its model config,
+           # it is important to set `use_cache=True` explicitly in the `generate` function
+           # to obtain optimal performance with BigDL-LLM INT4 optimizations
+
+           # Note that phi-1_5 uses GenerationConfig to enable 'use_cache'
+           output = model.generate(input_ids, do_sample=False, max_new_tokens=args.n_predict, generation_config = generation_config)
+           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, 'Prompt', '-'*20)
+           print(prompt)
+           print('-'*20, 'Output', '-'*20)
+           print(output_str)
+   ```
+   Here is the sample output on the laptop. 
+   ```
+   Inference time: 3.526491641998291 s
+   -------------------- Prompt --------------------
+   Question:What is AI?
+
+   Answer:
+   -------------------- Output --------------------
+   Question:What is AI?
+
+   Answer: AI stands for Artificial Intelligence, which is the simulation of human intelligence in machines.
+   ```
+
+
+