Deploy Meta’s LLaMA 3 on JD Cloud: A Complete Step‑by‑Step Tutorial

On April 19, Meta announced its latest large language model, LLaMA 3, offering an 8‑billion‑parameter and a 70‑billion‑parameter version with an 8K context window, touted as the strongest open‑source LLM to date and rivaling GPT‑4 in many benchmarks. Detailed evaluation results are available at the official report .

LLaMA 3 adopts a standard decoder‑only Transformer architecture, with performance gains largely attributed to higher‑quality data: 15 terabytes of pre‑training data (seven times that of LLaMA 2) and a significantly larger proportion of code to boost reasoning. The tokenizer’s vocabulary was expanded to 128 K tokens (up from 32 K), improving tokenization granularity, and both model sizes employ Grouped‑Query Attention (GQA) for faster inference.

The open‑source community reacted quickly; within five days, over a thousand variants appeared on Hugging Face and the number continues to grow.

Inspired by the AI wave, the author attempted to try LLaMA 3 via the official demo site but faced network delays, and a local installation was discouraged by the model’s 60 GB size and the high cost of GPU resources.

Turning to JD Cloud, the author outlines a practical deployment process:

Step 1

Enter the JD Cloud AI console at https://gcs-console.jdcloud.com/instance/list .

Step 2

Create a GPU instance, selecting the “pay‑by‑configuration” billing mode (≈ ¥1.89 per hour). Recharge ¥2 to obtain roughly two hours of compute.

Step 3

Wait for the instance status to become “Running”, then launch Jupyter from the console.

Step 4

Open a terminal in Jupyter and execute:

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

Step 5

In the file explorer, open llama-factory/src/web_demo.py, change server_port to 28888, and save.

Step 6

Back in the terminal, run the following commands:

cd /data/llama-factory</code>
<code>conda create -n liandan python=3.10 -y</code>
<code>conda activate liandan</code>
<code>pip install -e .[metrics]</code>
<code>CUDA_VISIBLE_DEVICES=0 python src/web_demo.py --model_name_or_path /gcs-pub/Meta-Llama-3-8B-Instruct --template llama3

Step 7

After a few minutes of fast inference, the model is accessible. Finally, in the JD Cloud console, select the instance, choose “Application → Custom Application”, and launch the LLaMA 3 prototype.

The platform also promises no‑code text‑to‑image capabilities, which the author plans to explore next.