GR00T-N1.7 — LIBERO object (GGUF for vla.cpp)
GGUF conversion of nvidia/GR00T-N1.7-LIBERO
for inference with vla.cpp, a lightweight
C++ inference engine for Vision-Language-Action models built on top of
llama.cpp.
GR00T-N1.7 pairs a Cosmos-Reason2 / Qwen3-VL-2B backbone (24-layer ViT + deepstack) with a hybrid action head (4-layer VL self-attention ⊕ 32-layer AlternateVLDiT flow-matching), producing 40×132 relative actions with q01/q99 normalisation. The vision tower is baked into the combined GGUF, so no separate mmproj file is needed.
Files
| File | Size | Description |
|---|---|---|
gr00t-n1d7-libero-object.gguf | 5.86 GiB | Combined VLA model — Qwen3-VL backbone + deepstack + VL self-attention + AlternateVLDiT action head + arch config, BF16 |
dataset_statistics.json | — | Action/state normalisation stats (required by the client) |
Usage
# Terminal 1 — serve (use the CUDA build for inference). No mmproj argument.
VLA_GR00T_BF16_WEIGHTS=1 VLA_GR00T_EMBODIMENT=libero_sim \
./build-cuda/vla-server --bind tcp://*:5566 \
gr00t-n1d7-libero-object.gguf
# Terminal 2 — drive a LIBERO episode (inside the LIBERO uv venv)
python eval/client/run_sim_client_direct.py \
--arch gr00t_n1_7 \
--task libero_object --task-id 0 --n-episodes 10 \
--stats-json dataset_statistics.json \
--vla-addr tcp://localhost:5566
Notes:
- Set
VLA_GR00T_EMBODIMENT=libero_simandVLA_GR00T_BF16_WEIGHTS=1(the latter is needed to fit an 8 GB card). - Pass
--stats-json dataset_statistics.json(action/state un-normalisation). - The client uses
--n-action-steps 16for this checkpoint.
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) | 16 | 98.0% | 10.26 ms | 164 ms | 6302 MiB VRAM |
| Jetson AGX Orin (sm_87) | 16 | 98.5% | 26.84 ms | 429 ms | 1317 MiB RAM |
GR00T-N1.7's ~6 GiB all-resident weight footprint OOMs the Jetson Orin Nano 8 GB unified pool, so it is not in the Orin Nano sweep. A
VLA_GR00T_STAGE_SWAP=1mode brings weights to ~3.25 GiB at a heavy per-call latency cost.
Implementation note
The action head's vlln must be fed the pre-norm LM hidden state
(hidden_states[-1], std ≈ 33), not the post-norm last_hidden_state (std ≈
2.17). vla.cpp drops lm_output_norm between the LM and the projector to match
HF's Qwen3VLBackbone.forward — this is the single largest contributor to going
from 0/10 to 10/10 on LIBERO.
License
Weights follow the upstream license of
nvidia/GR00T-N1.7-LIBERO
(NVIDIA license — review and accept it before use). The vla.cpp conversion
tooling and inference engine are MIT-licensed.