FlashDrive

Flash Vision-Language-Action Inference for Autonomous Driving

FlashDrive accelerates Alpamayo 1.5 and Alpamayo 1 — NVIDIA's 10B-parameter vision-language-action models for autonomous driving — by 4.5× with no loss in accuracy, through algorithm-system co-design of five techniques, none sufficient alone:

• Streaming inference — reuse each frame's KV cache across windows, so a new window only prefills its latest frames.
• Speculative reasoning (DFlash) — a block-diffusion draft proposes 8-token blocks that the target verifies in one forward, preserving its output distribution.
• W4A8 quantization (ParoQuant) — pairwise-rotation INT4 weights and INT8 activations via vLLM's Marlin kernels; the action expert stays bf16.
• Adaptive action caching — reuse the predicted velocity on select diffusion steps to skip action-expert forwards.
• System optimizations — static KV cache, fused expert projections, and torch.compile with CUDA graphs over the full decode loop.

Performance

Measured with scripts/infer.py on a single RTX PRO 6000 over 100 PhysicalAI-AV clips (one trajectory sample per step). minADE improves on both models.

Model	minADE ↓	Latency (per window)	Speedup
Alpamayo 1.5	1.705	717 ms	1.0×
Alpamayo 1.5 + FlashDrive	1.573	151 ms	4.7×
Alpamayo 1 (R1)	1.869	704 ms	1.0×
Alpamayo 1 (R1) + FlashDrive	1.662	155 ms	4.5×

Every stage of the rollout is accelerated (baseline → FlashDrive, ms):

Stage	Alpamayo 1.5	Alpamayo 1 (R1)
Encode	87.0 → 12.0 (7.3×)	86.4 → 12.0 (7.2×)
Prefill	165.3 → 47.2 (3.5×)	162.3 → 46.9 (3.5×)
Decode	271.7 → 45.3 (6.0×)	262.5 → 49.6 (5.3×)
Action	192.9 → 46.9 (4.1×)	192.6 → 46.9 (4.1×)

Quick Start

Installation

Clone the repo and install with uv (CUDA 12.8, Python 3.12, NVIDIA GPU with compute capability 8.0+):

git clone https://github.com/z-lab/flashdrive && cd flashdrive

uv venv --python 3.12
source .venv/bin/activate

uv sync

Inference

Benchmark a PhysicalAI-AV clip — reports minADE and per-window latency:

python scripts/infer.py --model-path z-lab/Alpamayo-1.5-10B    # optimized
python scripts/infer.py --model-path nvidia/Alpamayo-1.5-10B   # baseline

A z-lab checkpoint runs the optimized stack; an nvidia checkpoint runs the original model.

Library

import flashdrive

model = flashdrive.from_pretrained("z-lab/Alpamayo-1.5-10B")

pred_xyz, pred_rot = model.sample_trajectories_streaming(data)

The first call per stream only prefills the KV cache and returns (None, None); every later window returns trajectories. Passing an upstream nvidia/... checkpoint to flashdrive.from_pretrained loads the original model instead. See scripts/infer.py for building data from a PhysicalAI-AV clip.

Models

All checkpoints are on the Hugging Face collection. flashdrive.from_pretrained takes the base path; the W4A8 (-PARO) and DFlash (-DFlash) companions are derived from it by suffix and fetched automatically.

Model	Base	W4A8 (ParoQuant)	Draft (DFlash)
Alpamayo 1.5	z-lab/Alpamayo-1.5-10B	-PARO	-DFlash
Alpamayo 1 (R1)	z-lab/Alpamayo-R1-10B	-PARO	-DFlash

License

FlashDrive's code is released under the MIT License. The Alpamayo weights mirror NVIDIA's release and remain under its non-commercial license.

Citation

@article{li2026flashdrive,
  title  = {{FlashDrive: Flash Vision-Language-Action Inference for Autonomous Driving}},
  author = {Li, Zekai and Liang, Yihao and Zhang, Hongfei and Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  year   = {2026}
}

FlashDrive builds on DFlash and ParoQuant; please consider citing them as well:

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}

@inproceedings{liang2026paroquant,
  title     = {{ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference}},
  author    = {Liang, Yesheng and Chen, Haisheng and Zhang, Zihan and Han, Song and Liu, Zhijian},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}