llama2-chat-70b-4bit

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64ccdc322e592905f922a06e/VhwQtaklohkUXFWkjA-3M.png" width="450"/>

English | 简体中文

</div> <p align="center"> 👋 join us on <a href="https://twitter.com/intern_lm" target="_blank">Twitter</a>, <a href="https://discord.gg/xa29JuW87d" target="_blank">Discord</a> and <a href="https://r.vansin.top/?r=internwx" target="_blank">WeChat</a> </p>

W4A16 LLM Model Deployment

LMDeploy supports LLM model inference of 4-bit weight, with the minimum requirement for NVIDIA graphics cards being sm80.

Before proceeding with the inference, please ensure that lmdeploy(>=v0.0.14) is installed.

pip install 'lmdeploy>=0.0.14'

4-bit LLM model Inference

You can download the pre-quantized 4-bit weight models from LMDeploy's model zoo and conduct inference using the following command.

Alternatively, you can quantize 16-bit weights to 4-bit weights following the "4-bit Weight Quantization" section, and then perform inference as per the below instructions.

Take the 4-bit Llama-2-70B model from the model zoo as an example:

git-lfs install
git clone https://huggingface.co/lmdeploy/llama2-chat-70b-4bit

As demonstrated in the command below, first convert the model's layout using turbomind.deploy, and then you can interact with the AI assistant in the terminal

## Convert the model's layout and store it in the default path, ./workspace.
lmdeploy convert \
    --model-name llama2 \
    --model-path ./llama2-chat-70b-w4 \
    --model-format awq \
    --group-size 128

## inference
lmdeploy chat ./workspace

Serve with gradio

If you wish to interact with the model via web ui, please initiate the gradio server as indicated below:

lmdeploy serve gradio ./workspace --server_name {ip_addr} --server_port {port}

Subsequently, you can open the website http://{ip_addr}:{port} in your browser and interact with the model

Inference Performance

We benchmarked the Llama 2 7B and 13B with 4-bit quantization on NVIDIA GeForce RTX 4090 using profile_generation.py. And we measure the token generation throughput (tokens/s) by setting a single prompt token and generating 512 tokens. All the results are measured for single batch inference.

pip install nvidia-ml-py

python profile_generation.py \
 --model-path /path/to/your/model \
 --concurrency 1 8 --prompt-tokens 0 512 --completion-tokens 2048 512

4-bit Weight Quantization

It includes two steps:

• generate quantization parameter
• quantize model according to the parameter

Step 1: Generate Quantization Parameter

lmdeploy lite calibrate \
  --model $HF_MODEL \
  --calib_dataset 'c4' \             # Calibration dataset, supports c4, ptb, wikitext2, pileval
  --calib_samples 128 \              # Number of samples in the calibration set, if memory is insufficient, you can appropriately reduce this
  --calib_seqlen 2048 \              # Length of a single piece of text, if memory is insufficient, you can appropriately reduce this
  --work_dir $WORK_DIR \             # Folder storing Pytorch format quantization statistics parameters and post-quantization weight

Step2: Quantize Weights

LMDeploy employs AWQ algorithm for model weight quantization.

lmdeploy lite auto_awq \
  --model $HF_MODEL \
  --w_bits 4 \                       # Bit number for weight quantization
  --w_sym False \                    # Whether to use symmetric quantization for weights
  --w_group_size 128 \               # Group size for weight quantization statistics
  --work_dir $WORK_DIR \             # Directory saving quantization parameters from Step 1

After the quantization is complete, the quantized model is saved to $WORK_DIR. Then you can proceed with model inference according to the instructions in the "4-Bit Weight Model Inference" section.