---

Page | Paper | Model

News

• [2026/05] 🔥 Realtime-VLA FLASH code is now available.

Highlights

Realtime-VLA FLASH is the first speculative inference framework for diffusion-based VLAs.

• Speculative inference as fast as 7.8 ms (2 views), enabling over 125 Hz real-time inference.
• VLM-aligned draft architecture with a deployment-friendly block design.
• FLASH serving with customized Triton kernels, achieving a 3.04× average task-level speedup.

Installation

Follow openpi README:

git clone --recurse-submodules https://github.com/dexmal/realtime-vla-flash
# Or if you already cloned the repo:
git submodule update --init --recursive

Install the Python environment with uv:

GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

LIBERO client/evaluation code can run in a separate environment. (see LIBERO README).

Quick Start

First, convert the pretrained pi0 and draft checkpoints into the Triton weight layout.

uv run scripts/spec/triton/convert_for_triton.py \
   --mode base \
   --jax-path /path/to/jax/checkpoint \
   --output converted/base

uv run scripts/spec/triton/convert_for_triton.py \
   --mode draft \
   --draft-ckpt /path/to/draft_model.pt \
   --output converted/draft

Then start the policy server and the LIBERO client.

uv run scripts/spec/spec_serve_policy.py \
  --config pi0_libero \
  --base-triton-path converted/base \
  --draft-triton-path converted/draft \
  --task-suite-name libero_goal \
  --backend triton

uv run scripts/spec/spec_client_libero.py \
  --task-suite-name libero_goal

Benchmark

You can check the inference time on your local machine by

uv run python scripts/spec/pi0_benchmark.py

Train Draft Model

  uv run scripts/spec/enc_cache.py \
    --config pi0_libero \
    --checkpoint-dir /openpi-assets/checkpoints/pi0_libero_torch \
    --task-suite-name libero_goal \
    --output-dir /tmp/spec_quickstart_train/libero_goal_cache

  uv run scripts/spec/spec_draft_train.py \
    --cache-dir /tmp/spec_quickstart_train/libero_goal_cache \
    --output draft_model_goal_torch.pt

A typical workflow is:

1. Build a prefix-embedding cache with scripts/spec/enc_cache.py.
2. Train the draft head with scripts/spec/spec_draft_train.py.
3. Serve the FLASH policy with scripts/spec/spec_serve_policy.py.
4. Run LIBERO client evaluation or sweeps with scripts/spec/spec_client_libero.py or scripts/spec/exp/run_sweep.py.

Citation

If you find this work useful, please cite the paper once the arXiv version is available:

@article{niu2026realtimevlaflash,
  title={Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs},
  author={Niu, Jiahui and Gu, Kefan and Zhao, Yucheng and Liang, Shengwen and Wang, Tiancai and Hu, Xing and Wang, Ying and Li, Huawei},
  journal={arXiv preprint arXiv:2605.13778},
  year={2026}
}

Acknowledgements

• dexmal/realtime-vla
• openpi