Skip to content
Blueprint

← All articles · Deploy

Installing the runtime

The runtime is llama.cpp — Blueprint downloads the pre-built binary for your OS and verifies it before letting you serve anything.

2 min read

What "runtime" means here

Blueprint's default inference engine is llama.cpp's llama-server — the same binary you'd build yourself by cloning the ggerganov/llama.cpp repo and running make. It's a single C++ executable that loads GGUF model files and serves an OpenAI-compatible HTTP API on a port you choose.

Blueprint downloads a pre-built release from llama.cpp's GitHub releases page — no compilation on your machine, no CUDA toolkit required. The right build is picked for your OS + GPU vendor:

  • Windows + NVIDIA: CUDA-enabled build
  • Linux + NVIDIA: CUDA-enabled build
  • Linux + AMD: ROCm build
  • macOS Apple Silicon: Metal-enabled build
  • macOS Intel: CPU-only build

The runtime lives at ~/.blueprint/runtime/llama-server[.exe]. Updates happen on demand from the Maintain tab.

Why we ship llama.cpp specifically

Among the open-engine choices (llama.cpp, vLLM, TensorRT-LLM, exllama, MLC, etc.), llama.cpp is the broadest-compatibility option:

  • Runs anywhere: CPU, NVIDIA, AMD, Apple, even on a Raspberry Pi
  • No Python dependency: single binary
  • GGUF support: the most published quant format in the open ecosystem
  • Reasonable throughput: not the fastest, but within 30% of vLLM on single-GPU workloads

vLLM and TensorRT-LLM are stubbed in the Engines card for future implementation — they win on throughput at the cost of complexity (Python, much heavier install).

How to install

  1. Open Dashboard → Maintain → Runtime.
  2. Click Install runtime.
  3. Blueprint downloads the right build (~150-300 MB depending on OS), extracts it, and verifies the binary launches.

The install is idempotent — running it again upgrades to the latest release if your version is behind.

Verifying the install

After install, the Maintain tab's Runtime card shows:

  • Installed version (e.g. b4321) — matches llama.cpp's release tag
  • Latest available — what's on GitHub right now
  • Binary path — for sanity-check via terminal

If installed === latest, you're current. If not, the Update button pulls the new release.

Common pitfalls

  • CUDA driver mismatch on Linux/Windows: the prebuilt CUDA llama builds expect CUDA 12.x runtime. If your driver is older, the binary loads but reports "no CUDA device" and falls back to CPU. Fix: update your NVIDIA driver, not the runtime.
  • Antivirus quarantining the download: some EDR / corporate AV flags newly-downloaded EXEs. If the install fails silently on Windows, check your AV's quarantine folder.
  • Disk space: the runtime is small (~300 MB) but each model you pull is multiple GB. Watch free space before pulling.

Need help with this in production?

Inspire AI Lab runs LLM optimization engagements end-to-end using Blueprint. If you'd rather hand the work to us instead of running it yourself, book a 30-minute review.