Deploying Meta LLaMA 3 on JD Cloud: A Step‑by‑Step Tutorial

On April 19, Meta announced the open‑source LLaMA 3 family, offering an 8‑billion‑parameter and a 70‑billion‑parameter model with an 8K context window, claiming performance that rivals GPT‑4 and surpasses many proprietary models in benchmark tests.

The models use a standard decoder‑only Transformer architecture, benefit from a seven‑fold increase in pre‑training data (15 TB) and a larger 128K tokenizer, and employ Grouped‑Query Attention (GQA) to improve inference efficiency.

The open‑source community responded quickly, with over a thousand variants appearing on Hugging Face within five days.

The author attempted to access LLaMA 3 via Meta's online demo but faced network delays, then decided to run the model locally on JD Cloud due to the large model size (≈60 GB) and GPU requirements.

Step 1: Log into the JD Cloud AI console at https://gcs-console.jdcloud.com/instance/list .

Step 2: Create a GPU instance, selecting the “按配置” (pay‑by‑configuration) billing option; the hourly price is ¥1.89, so a ¥2 recharge yields about two hours of usage.

Step 3: Wait for the instance status to become “Running”, then click the Jupyter button to open the AI development environment.

Step 4: In the Jupyter terminal, copy the factory code to the data directory:

cp -r /gcs-pub/llama-factory/ /data/

Step 5: Open llama-factory/src/web_demo.py in the file explorer and change server_port to 28888 , then save.

Step 6: In the terminal, run the following commands to set up the environment and launch the demo:

cd /data/llama-factory

conda create -n liandan python=3.10 -y

conda activate liandan

pip install -e .[metrics]

CUDA_VISIBLE_DEVICES=0 python src/web_demo.py --model_name_or_path /gcs-pub/Meta-Llama-3-8B-Instruct --template llama3

The platform is noted for its fast inference speed, allowing the model to respond within minutes.

Step 7: Return to the JD Cloud console, select the instance, choose “应用 → 自定义应用”, and the LLaMA 3 prototype becomes accessible.

The author concludes that the platform also supports no‑code text‑to‑image applications and expresses excitement to experiment further with LLaMA 3.