BitVLA — LIBERO object (GGUF for vla.cpp)

GGUF conversion of hongyuw/ft-bitvla-bitsiglipL-224px-libero_object-bf16 for inference with vla.cpp, a lightweight C++ inference engine for Vision-Language-Action models built on top of llama.cpp.

BitVLA is a 1.58-bit ternary VLA: BitSigLIP-L vision tower + a 30-layer BitNet language model (ternary weights, squared_relu, attn/ffn sub-norm) + a bidirectional-attention MLPResNet regression head (not flow matching). Its vision tower is baked into the combined GGUF, so no separate mmproj file is needed.

On the full libero_object sweep BitVLA scores 100%, beating the paper's 99.6%.

Files

File	Size	Description
`bitvla-libero-object.gguf`	5.39 GiB	Combined VLA model, BF16 — portable / CPU-runnable
`bitvla-libero-object-int2.gguf`	1.34 GiB	Combined VLA model, int2-packed — CUDA-only; ~1.15 GiB host-RAM peak at load (≈10× smaller than the BF16 build), ideal for the 8 GB Jetson target

Pick one GGUF. The int2-packed build is byte-for-byte equivalent in behaviour and is the recommended choice for GPU / Jetson deployment.

Usage

# Terminal 1 — serve (use the CUDA build for inference). No mmproj argument.
./build-cuda/vla-server --bind tcp://*:5566 \
    bitvla-libero-object-int2.gguf

# Terminal 2 — drive a LIBERO episode (inside the LIBERO uv venv)
python eval/client/run_sim_client_direct.py \
    --arch bitvla \
    --task libero_object --task-id 0 --n-episodes 10 \
    --n-action-steps 8 --max-length 600 \
    --vla-addr tcp://localhost:5566

Notes:

Pass --max-length 600.
--tokenizer is optional — the client auto-loads the chat-template tokenizer and dataset_statistics.json from the upstream Hub repo. Pass `--tokenizer` only to override with a local/offline copy.

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	100.0%	37.85 ms	303 ms	1312 MiB VRAM
Jetson AGX Orin (sm_87)	8	100.0%	101.11 ms	809 ms	1149 MiB RAM
Jetson Orin Nano 8 GB (sm_87)	8	100.0%	355.65 ms	2845 ms	2199 MiB RAM

Per-step latency is reported after the IMMA tensor-core ladder-GEMM rework: the 30-layer BitNet LM prefill dropped from 999 → 146 ms (6.85×) on an RTX 5070, and end-to-end env-step latency fell 172.8 → 37.85 ms (4.6×) on the RTX 3060 with behaviour unchanged (still 100%).

License

Weights follow the upstream license of hongyuw/ft-bitvla-bitsiglipL-224px-libero_object-bf16. The vla.cpp conversion tooling and inference engine are MIT-licensed.