SmolVLA - LIBERO (GGUF for vla.cpp)

GGUF conversion of HuggingFaceVLA/smolvla_libero for inference with vla.cpp, a lightweight C++ inference engine for Vision-Language-Action models built on top of llama.cpp.

SmolVLA-450M is the golden reference architecture in vla.cpp - a SigLIP vision tower + SmolLM2 language backbone + flow-matching action expert. Every other arch in the engine is validated against the SmolVLA recipe.

Usage

Build vla-server from the vla.cpp repo, then:

# Terminal 1 - serve (use the CUDA build for inference)
./build-cuda/vla-server --bind tcp://*:5566 \
    smolvla-libero.gguf

# Terminal 2 - drive a LIBERO episode (inside the LIBERO uv venv)
python eval/client/run_sim_client_direct.py \
    --arch smolvla \
    --task libero_object --task-id 0 --n-episodes 10 \
    --vla-addr tcp://localhost:5566

smolvla uses --n-action-steps 1 (the client default).

Benchmark

Full libero_object sweep (10 tasks × 20 episodes = 200 episodes), vla-server + run_sim_client_direct.py:

Hardware	n_act	Success rate	client/step	client/call	Peak mem
RTX 3060 (sm_86)	4	90.5%	28.16 ms	113 ms	1410 MiB VRAM
Jetson AGX Orin (sm_87)	4	90.5%	65.41 ms	262 ms	689 MiB RAM
Jetson Orin Nano 8 GB (sm_87)	4	88.0%	141.81 ms	567 ms	2031 MiB RAM

License

Weights follow the upstream license of HuggingFaceVLA/smolvla_libero (Apache-2.0). The vla.cpp conversion tooling and inference engine are MIT-licensed.