dnf install epol-release

# Install open source modules
dnf install -y kmod-nvidia-570-dkms-open nvidia-570 nvidia-driver-570-open
 
# Install closed-source modules
dnf install -y kmod-nvidia-570-dkms nvidia-570 nvidia-driver-570

# Restart the machine
sudo reboot
# View installed driver software packages
dnf list installed | grep nvidia
# View driver information
nvidia-smi

# Uninstall open source modules
dnf remove -y kmod-nvidia-570-dkms-open nvidia-570 nvidia-driver-570-open
 
# Uninstall closed-source modules
dnf remove -y kmod-nvidia-570-dkms nvidia-570 nvidia-driver-570

# View installed cuda software packages
dnf list installed | grep cuda
# View the cuda version
nvcc --version

# Download uv 
dnf install pip
pip install uv
 
# Recommended: Use uv for installation to intelligently resolve dependency conflicts.
uv pip install vllm==0.10.0 --system

# Start the vllm service and pull a large model (Qwen is used as an example here).
vllm serve "Qwen/Qwen3-0.6B" 
 
# If the machine has a low configuration, try the following command to start the service.
vllm serve "Qwen/Qwen3-0.6B" --gpu-memory-utilization 0.7 --max-model-len 1024 --max-num-seqs 4 --dtype float16 --block-size 16

# Open a new terminal
curl -X POST "http://localhost:8000/v1/chat/completions" -H "Content-Type: application/json" --data '{"model": "Qwen/Qwen3-0.6B","messages": [{"role": "user","content": "What is the capital of France?"}]}'

Create a python program with the following content:
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
 
# Initialize model parameters.
model_path = "Qwen/Qwen3-0.6B"  # Replace with the actual model path.
llm = LLM(
    model=model_path,
    dtype="half",  # Enable half-precision inference (FP16).
    trust_remote_code=True,  # Required for Hugging Face models.
    tensor_parallel_size=1,  # Run on a single GPU.
    gpu_memory_utilization=0.8  # GPU memory utilization control.
)
 
# Configure generation parameters.
sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.95,
    max_tokens=256,
    repetition_penalty=1.1
)
 
# Build a conversation template.
messages = [
    {"role": "system", "content": "You are a poet"},
    {"role": "user", "content": "Write a seven-character quatrain about spring."}
]
tokenizer = AutoTokenizer.from_pretrained(model_path)
prompt = tokenizer.apply_chat_template(
    conversation=messages,
    tokenize=False,
    add_generation_prompt=True
)
 
# Run inference.
outputs = llm.generate(
    prompts=[prompt],
    sampling_params=sampling_params
)
 
# Output results.
for output in outputs:
    generated_text = output.outputs[0].text
    print(f"Generated result:\\n{generated_text}")

# Install dependencies.
dnf install -y gawk
# Download and run the installation script.
curl -fsSL https://ollama.com/install.sh | sh

# Start the service.
ollama serve

# Pull a large model (Qwen is used as an example here).
ollama run qwen3:0.6b

export LD_LIBRARY_PATH=/usr/local/lib/python3.11/site-packages/nvidia/cuda_cupti/lib:$LD_LIBRARY_PATH