Usage: bash run.sh [-c <llm_api>] [-i <device_id>] [-b <runtime_backend>] [-m <model_name>] [-t <conv_template>] [-p <tensor_parallel>] [-r <gpu_memory_utilization>]

-c <llm_api>: "Options {local, cloud} to specify the llm API mode, default is 'local'. If set to '-c cloud', please mannually set the environments {OPENAI_API_KEY, OPENAI_API_BASE, OPENAI_API_MODEL_NAME, OPENAI_API_CONTEXT_LENGTH} into .env fisrt."
-i <device_id>: "Specify argument GPU device_id."
-b <runtime_backend>: "Specify argument LLM inference runtime backend, options={default, hf, vllm}"
-m <model_name>: "Specify argument the model name to load public LLM model using FastChat serve API, options={Qwen-7B-Chat, deepseek-llm-7b-chat, ...}"
-t <conv_template>: "Specify argument the conversation template according to the public LLM model when using FastChat serve API, options={qwen-7b-chat, deepseek-chat, ...}"
-p <tensor_parallel>: "Use options {1, 2} to set tensor parallel parameters for vllm backend when using FastChat serve API, default tensor_parallel=1"
-r <gpu_memory_utilization>: "Specify argument gpu_memory_utilization (0,1] for vllm backend when using FastChat serve API, default gpu_memory_utilization=0.81"
-h: "Display help usage message"

Note: You can choose the most suitable Service Startup Command based on your own device conditions.
(1) Local Embedding/Rerank will run on device gpu_id_1 when setting "-i 0,1", otherwise using gpu_id_0 as default.
(2) When setting "-c cloud" that will use local Embedding/Rerank and OpenAI LLM API, which only requires about 4GB VRAM (recommend for GPU device VRAM <= 8GB).
(3) When you use OpenAI LLM API, you will be required to enter {OPENAI_API_KEY, OPENAI_API_BASE, OPENAI_API_MODEL_NAME, OPENAI_API_CONTEXT_LENGTH} immediately.
(4) "-b hf" is the most recommended way for running public LLM inference for its compatibility but with poor performance.
(5) When you choose a public Chat LLM for QAnything system, you should take care of a more suitable **PROMPT_TEMPLATE** (/path/to/QAnything/qanything_kernel/configs/model_config.py) setting considering different LLM models.

## Step 1. Download the public LLM model (e.g., Qwen-7B-QAnything) and save to "/path/to/QAnything/assets/custom_models"
## (Optional) Download Qwen-7B-QAnything from ModelScope: https://www.modelscope.cn/models/netease-youdao/Qwen-7B-QAnything
## (Optional) Download Qwen-7B-QAnything from Hugging Face: https://huggingface.co/netease-youdao/Qwen-7B-QAnything
cd /path/to/QAnything/assets/custom_models
git clone https://huggingface.co/netease-youdao/Qwen-7B-QAnything

## Step 2. Execute the service startup command.  Here we use "-b hf" to specify the Hugging Face transformers backend.
## Here we use "-b hf" to specify the transformers backend that will load model in 8 bits but do bf16 inference as default for saving VRAM.
cd /path/to/QAnything
bash ./run.sh -c local -i 0 -b hf -m Qwen-7B-QAnything -t qwen-7b-qanything

## Step 1. Download the public LLM model (e.g., MiniChat-2-3B) and save to "/path/to/QAnything/assets/custom_models"
cd /path/to/QAnything/assets/custom_models
git lfs clone https://huggingface.co/GeneZC/MiniChat-2-3B

## Step 2. Execute the service startup command.  Here we use "-b hf" to specify the Hugging Face transformers backend.
## Here we use "-b hf" to specify the transformers backend that will load model in 8 bits but do bf16 inference as default for saving VRAM.
cd /path/to/QAnything
bash ./run.sh -c local -i 0 -b hf -m MiniChat-2-3B -t minichat

## Step 1. Download the public LLM model (e.g., Qwen-7B-QAnything) and save to "/path/to/QAnything/assets/custom_models"
## (Optional) Download Qwen-7B-QAnything from ModelScope: https://www.modelscope.cn/models/netease-youdao/Qwen-7B-QAnything
## (Optional) Download Qwen-7B-QAnything from Hugging Face: https://huggingface.co/netease-youdao/Qwen-7B-QAnything
cd /path/to/QAnything/assets/custom_models
git clone https://huggingface.co/netease-youdao/Qwen-7B-QAnything

## Step 2. Execute the service startup command.  Here we use "-b vllm" to specify the vllm backend.
## Here we use "-b vllm" to specify the vllm backend that will do bf16 inference as default.
## Note you should adjust the gpu_memory_utilization yourself according to the model size to avoid out of memory (e.g., gpu_memory_utilization=0.81 is set default for 7B. Here, gpu_memory_utilization is set to 0.85 by "-r 0.85").
cd /path/to/QAnything
bash ./run.sh -c local -i 0 -b vllm -m Qwen-7B-QAnything -t qwen-7b-qanything -p 1 -r 0.85

## Step 1. Download the public LLM model (e.g., MiniChat-2-3B) and save to "/path/to/QAnything/assets/custom_models"
cd /path/to/QAnything/assets/custom_models
git clone https://huggingface.co/GeneZC/MiniChat-2-3B

## Step 2. Execute the service startup command. 
## Here we use "-b vllm" to specify the vllm backend that will do bf16 inference as default.
## Note you should adjust the gpu_memory_utilization yourself according to the model size to avoid out of memory (e.g., gpu_memory_utilization=0.81 is set default for 7B. Here, gpu_memory_utilization is set to 0.5 by "-r 0.5").
cd /path/to/QAnything
bash ./run.sh -c local -i 0 -b vllm -m MiniChat-2-3B -t minichat -p 1 -r 0.5

## (Optional) Step 2. Execute the service startup command to specify the vllm backend by "-i 0,1 -p 2". It will do faster inference by setting a tensor parallel mode on 2 GPUs.
## bash ./run.sh -c local -i 0,1 -b vllm -m MiniChat-2-3B -t minichat -p 2 -r 0.5

## Trick 1. (Recommend for VRAM<=12 GB or GPU Compute Capability < 7.5) Using PaddleOCR serve in CPU mode **use_gpu=False** in '/path/to/QAnything/qanything_kernel/dependent_server/ocr_serve/ocr_server.py'
# GPU Compute Capability: https://developer.nvidia.com/cuda-gpus
# Note that **use_gpu=False** must be set when using RTX-1080Ti GPU, otherwise PaddleOCR will always return **empty ocr result** when using **use_gpu=True**.
ocr_engine = PaddleOCR(use_angle_cls=True, lang="ch", use_gpu=False, show_log=False)

## Trick 2. Try 1.8B/3B size LLM, such as Qwen-1.8B-Chat and MiniChat-2-3B.

## Trick 3. Try to limit the max length of context window by decreasing the value of **token_window** and increasing that of **offcut_token**
# /path/to/QAnything/qanything_kernel/connector/llm/llm_for_fastchat.py
# /path/to/QAnything/qanything_kernel/connector/llm/llm_for_local.py

## Trick 4. Try INT4-Weight-Only Quantization methods such as GPTQ/AWQ. You should take care of the sampling parameters considering possible loss of accuracy.