vrfai/InternVLA-N1-DualVLN-FP8

InternVLA-N1-DualVLN — FP8

FP8 quantization of InternRobotics/InternVLA-N1-DualVLN, for deployment on NVIDIA Jetson Thor with TensorRT Edge-LLM. This support is pending review as NVIDIA/TensorRT-Edge-LLM#193; until it merges, build from that PR's branch directly from the upstream repo:

git clone https://github.com/NVIDIA/TensorRT-Edge-LLM.git
cd TensorRT-Edge-LLM
git fetch origin pull/193/head:internvla-n1
git checkout internvla-n1

What's quantized

InternVLA-N1-DualVLN is a dual-system vision-language navigation model: a Qwen2.5-VL-7B planner (System 2) plus a flow-matching trajectory expert with a DINOv2 memory block (System 1), joined by a z_latents bridge. Only System 2 is quantized here. System 1 and the bridge (cond_projector, latent_queries) are kept at the source dtype — the bridge is four rows through a Linear/GELU/Linear, so quantizing it saves nothing measurable and would put error directly on the tensor System 1 steers by.

Produced with tensorrt-edgellm-quantize llm --quantization fp8, verified end to end: model_type reads internvla_n1, and cond_projector / latent_queries / traj_dit / rgb_model are present and unquantized in the export.

Measured on Jetson Thor (sm_110, JetPack 7.1, TensorRT 10.13.3.9)

	prefill (1024 tok)	decode (pastKV 1024)	first plan	control rate	engine
PyTorch bf16	328.9 ms	99.4 ms	160 ms	208.1 ms (4.8 Hz)	~15 GB
This checkpoint	90.6 ms	32.8 ms	118 ms	61.3 ms (16.3 Hz)	7.10 GB

Closed-loop navigation success, 199 R2R val_unseen episodes (habitat-sim on Thor, VLN_TRAJ_SEED=100), against PyTorch bf16 (69.8% SR):

SR	agrees with PyTorch	McNemar p
68.3%	166/199	0.728

Not distinguishable from PyTorch at this sample size (p = 0.728). See the PR that added this support for the full validation, including why offline metrics like bridge/trajectory cosine were tried and rejected as acceptance criteria for this model — closed-loop SR is the only thing that agreed with itself across quantization schemes.

Usage

Export and build with TensorRT Edge-LLM:

tensorrt-edgellm-export /path/to/this/checkpoint ./onnx
export EDGELLM_PLUGIN_PATH=.../libNvInfer_edgellm_plugin.so

build/examples/llm/llm_build --onnxDir onnx/llm --engineDir engines/llm \
    --maxBatchSize 1 --maxInputLen 3072 --maxKVCacheCapacity 4096

System 1 is not part of this checkpoint's quantization and is built separately — see the model's runtime guide, experimental_models/internvla_n1/README.md in the checked-out branch above (or on PR #193 once merged) for the full export → build → run flow (both systems, asynchronous, one process).

License

Inherits the source model's license, CC BY-NC-SA 4.0. Non-commercial use, share-alike, attribution required. See InternRobotics/InternVLA-N1-DualVLN for the full terms.

Citation

If you use this checkpoint, please cite the original model:

@misc{internvla-n1,
    title = {{InternVLA-N1: An} Open Dual-System Navigation Foundation Model with Learned Latent Plans},
    author = {InternVLA-N1 Team},
    year = {2025},
    booktitle={arXiv},
}
@misc{wei2025groundslowfastdualsystem,
      title={Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and-Language Navigation},
      author={Meng Wei and Chenyang Wan and Jiaqi Peng and Xiqian Yu and Yuqiang Yang and Delin Feng and Wenzhe Cai and Chenming Zhu and Tai Wang and Jiangmiao Pang and Xihui Liu},
      year={2025},
      eprint={2512.08186},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.08186},
}