How to install vLLM with pip

What this does

Installs the vLLM inference engine into an existing Python environment, providing a pip package that enables high-throughput LLM serving on CUDA hardware. The result is a working vllm command-line entry point and Python API.

Steps

1. Create a clean virtual environment. Isolating vLLM prevents dependency conflicts with other packages.

   python -m venv vllm-env
   source vllm-env/bin/activate
   

   Expected output: the prompt shows (vllm-env).

2. Upgrade pip and setuptools. Older setuptools can cause wheels to fail during build.

   pip install --upgrade pip setuptools wheel
   

   Expected output: Successfully installed pip-X.Y.Z setuptools-X.Y.Z wheel-X.Y.Z.

3. Install vLLM stable release.

   pip install vllm
   

   Expected output: Successfully installed vllm-X.Y.Z. This step downloads pre-built CUDA wheels; expect 1-3 minutes on a fast connection.

4. Verify installation.

   python -c "import vllm; print(vllm.__version__)"
   

   Expected output: the version string, e.g. 0.8.5.

Verification

python -c "import vllm; print('vLLM version:', vllm.__version__)"
# Expected: vLLM version: 0.x.y

Common failures

• cublas / nccl not found — CUDA version mismatch. Ensure the CUDA toolkit version matches the driver. Run nvidia-smi first, then install vLLM specifying the correct CUDA version.
• torch version conflict — Another package pins an older PyTorch. Create a fresh virtual environment.
• Out-of-memory during extension build — CUDA kernel compilation requires ~4 GB free RAM per GPU. Close other GPU processes before installing.
• Permission denied on pip writes — Use --user flag or a virtual environment instead of system-wide install.
• pre-built wheel not available for this platform — Build from source or use the nightly wheel: pip install vllm --pre.

Related guides

• How to run vLLM with a HuggingFace model
• How to configure vLLM GPU memory allocation
• Course Local AI Fundamentals