10 Star 20 Fork 12

DeepSpark/DeepSparkInference

加入 Gitee
与超过 1200万 开发者一起发现、参与优秀开源项目,私有仓库也完全免费 :)
免费加入
文件
克隆/下载
贡献代码
同步代码
取消
提示: 由于 Git 不支持空文件夾,创建文件夹后会生成空的 .keep 文件
Loading...
README

ChatGLM3-6B (vLLM)

Model Description

ChatGLM3-6B is trained on large-scale natural language text data, enabling it to understand and generate text. It can be applied to various natural language processing tasks such as dialogue generation, text summarization, and language translation.

Supported Environments

Iluvatar GPU IXUCA SDK
MR-V100 4.2.0

Model Preparation

Prepare Resources

Pretrained model: https://huggingface.co/THUDM/chatglm3-6b

mkdir /data/chatglm/
mv chatglm3-6b.zip/tar /data/chatglm/

Install Dependencies

In order to run the model smoothly, you need to get the sdk from resource center of Iluvatar CoreX official website.

# Install libGL
## CentOS
yum install -y mesa-libGL
## Ubuntu
apt install -y libgl1-mesa-glx

pip3 install vllm
pip3 install transformers

Model Inference

python3 offline_inference.py --model /data/chatglm/chatglm3-6b --trust-remote-code --temperature 0.0 --max-tokens 256

Use the server

Start the server.

python3 -m vllm.entrypoints.openai.api_server --model /data/chatglm/chatglm3-6b --gpu-memory-utilization 0.9 --max-num-batched-tokens 8193 \
        --max-num-seqs 32 --disable-log-requests --host 127.0.0.1 --port 12345 --trust-remote-code

Test using the OpenAI interface

python3 server_inference.py --host 127.0.0.1 --port 12345 --model_path /data/chatglm/chatglm3-6b

Benchmarking vLLM

# Downloading the ShareGPT dataset.
wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json

# Cloning the vllm project
git clone https://github.com/vllm-project/vllm.git -b v0.5.4 --depth=1

Starting server.

python3 -m vllm.entrypoints.openai.api_server --model /data/chatglm/chatglm3-6b --gpu-memory-utilization 0.9 --max-num-batched-tokens 8193 \
        --max-num-seqs 32 --disable-log-requests --host 127.0.0.1 --trust-remote-code

Starting benchmark client.

python3 benchmark_serving.py --host 127.0.0.1 --num-prompts 16 --model /data/chatglm/chatglm3-6b --dataset-name sharegpt \
        --dataset-path /data/dataset/ShareGPT_V3_unfiltered_cleaned_split.json --sharegpt-output-len 130 --trust-remote-code
马建仓 AI 助手
尝试更多
代码解读
代码找茬
代码优化
Python
1
https://gitee.com/deep-spark/deepsparkinference.git
git@gitee.com:deep-spark/deepsparkinference.git
deep-spark
deepsparkinference
DeepSparkInference
master

搜索帮助

371d5123 14472233 46e8bd33 14472233