vrfai/evo1-libero-gguf

Evo-1 — LIBERO (GGUF for vla.cpp)

GGUF conversion of MINT-SJTU/Evo1_LIBERO for inference with vla.cpp, a lightweight C++ inference engine for Vision-Language-Action models built on top of llama.cpp.

Evo-1 couples an InternVL3-1B vision-language backbone with a cross-attention DiT flow-matching action head. Its vision tower is baked into the combined GGUF, so no separate mmproj file is needed.

Files

File	Size	Description
evo1-libero.gguf	1.45 GiB	Combined VLA model — InternVL3 LM + vision tower + cross-attn DiT action head + dataset stats + arch config, BF16

Usage

# Terminal 1 — serve (use the CUDA build for inference). No mmproj argument.
./build-cuda/vla-server --bind tcp://*:5566 \
    evo1-libero.gguf

# Terminal 2 — drive a LIBERO episode (inside the LIBERO uv venv)
python eval/client/run_sim_client_direct.py \
    --arch evo1 \
    --task libero_object --task-id 0 --n-episodes 10 \
    --vla-addr tcp://localhost:5566

Benchmark

Full libero_object sweep (10 tasks × 20 episodes = 200 episodes):

Hardware	n_act	Success rate	client/step	client/call	Peak mem
RTX 3060 (sm_86)	8	94.5%	63.60 ms	509 ms	1564 MiB VRAM
Jetson AGX Orin (sm_87)	8	95.5%	131.01 ms	1048 ms	638 MiB RAM
Jetson Orin Nano 8 GB (sm_87)	8	97.5%	458.84 ms	3671 ms	2135 MiB RAM

Implementation note

Evo-1 exposed a Qwen2 + flash-attention-2 masking subtlety: HF's FA2 path zeroes the attention output of masked queries, while a naïve softmax computes real attention for them — contaminating the LM context at image-context positions. vla.cpp mirrors HF exactly with a per-query mask, which is what takes this checkpoint from 0/5 to passing on LIBERO.

License

Weights follow the upstream license of MINT-SJTU/Evo1_LIBERO. The vla.cpp conversion tooling and inference engine are MIT-licensed.