High-standard model card (FlashDrive); NVIDIA non-commercial license

README.md

@@ -1,76 +1,69 @@
 ---
-pipeline_tag: robotics
-library_name: transformers
 license: other
+license_name: nvidia-license
+license_link: https://huggingface.co/nvidia/Alpamayo-1.5-10B/blob/main/LICENSE
+pipeline_tag: robotics
 language:
 - en
+inference: false
 base_model:
 - nvidia/Alpamayo-R1-10B
+new_version: z-lab/Alpamayo-1.5-10B
+datasets:
+- nvidia/PhysicalAI-Autonomous-Vehicles
+- nvidia/PhysicalAI-Autonomous-Vehicles-NuRec
 tags:
 - flashdrive
 - autonomous-driving
 - vision-language-action
 - alpamayo
-new_version: z-lab/Alpamayo-1.5-10B
 ---
 
-# Alpamayo-R1-10B (FlashDrive)
+# Alpamayo 1 (R1) (FlashDrive)
 
-[Blog](https://z-lab.ai/projects/flashdrive/) | [GitHub](https://github.com/z-lab/flashdrive)
+[GitHub](https://github.com/z-lab/flashdrive) | [Blog](https://z-lab.ai/projects/flashdrive/) | [Models](https://huggingface.co/collections/z-lab/flashdrive)
 
-**FlashDrive** is an algorithm-system co-design inference framework for
-[Alpamayo](https://huggingface.co/nvidia/Alpamayo-1.5-10B), NVIDIA's 10B-parameter
-vision-language-action (VLA) models for autonomous driving. It combines streaming VLM
-inference, [DFlash](https://github.com/z-lab/dflash) speculative reasoning, adaptive action
-caching, [ParoQuant](https://github.com/z-lab/paroquant) W4A8 quantization, and system
-optimizations to accelerate inference while preserving accuracy.
+[FlashDrive](https://github.com/z-lab/flashdrive) accelerates [Alpamayo 1 (R1)](https://huggingface.co/nvidia/Alpamayo-R1-10B) — one of NVIDIA's 10B-parameter vision-language-action models for autonomous driving — by **3.9× with no loss in accuracy**, through streaming inference, [DFlash](https://github.com/z-lab/dflash) speculative reasoning, [ParoQuant](https://github.com/z-lab/paroquant) W4A8 quantization, adaptive action caching, and `torch.compile`.
 
-This repository mirrors the weights of
-[nvidia/Alpamayo-R1-10B](https://huggingface.co/nvidia/Alpamayo-R1-10B) and serves as the
-**base checkpoint** of the FlashDrive stack. FlashDrive derives the companion checkpoints
-from this repo id by suffix:
+This repository mirrors the weights of [nvidia/Alpamayo-R1-10B](https://huggingface.co/nvidia/Alpamayo-R1-10B) and is the **base checkpoint** of the FlashDrive stack. Loading it pulls the derived companions automatically:
 
-| Repository | Contents |
+| Checkpoint | Contents |
 |---|---|
-| [z-lab/Alpamayo-R1-10B-PARO](https://huggingface.co/z-lab/Alpamayo-R1-10B-PARO) | W4A8 (ParoQuant) language-model weights |
-| [z-lab/Alpamayo-R1-10B-DFlash](https://huggingface.co/z-lab/Alpamayo-R1-10B-DFlash) | DFlash block-diffusion draft model |
+| [z-lab/Alpamayo-R1-10B-PARO](https://huggingface.co/z-lab/Alpamayo-R1-10B-PARO) | W4A8 ([ParoQuant](https://github.com/z-lab/paroquant)) language-model weights |
+| [z-lab/Alpamayo-R1-10B-DFlash](https://huggingface.co/z-lab/Alpamayo-R1-10B-DFlash) | [DFlash](https://github.com/z-lab/dflash) block-diffusion draft model |
 
-## Quick Start
+## Usage
 
-Install [FlashDrive](https://github.com/z-lab/flashdrive), then load the base checkpoint —
-the `-PARO` and `-DFlash` companions are derived from it by suffix and fetched automatically:
+Install [FlashDrive](https://github.com/z-lab/flashdrive), then load this base checkpoint — the `-PARO` and `-DFlash` companions are fetched automatically:
 
 ```python
-from flashdrive import load_flashdrive_model
+import flashdrive
 
-model = load_flashdrive_model("z-lab/Alpamayo-R1-10B")
+model = flashdrive.from_pretrained("z-lab/Alpamayo-R1-10B")
 
 pred_xyz, pred_rot = model.sample_trajectories_streaming(data)
 ```
 
-The first call per stream only prefills the KV cache and returns `(None, None)`.
-For an end-to-end benchmark on a PhysicalAI-AV clip:
+The first call per stream only prefills the KV cache and returns `(None, None)`; every later window returns trajectories. For an end-to-end benchmark on a PhysicalAI-AV clip:
 
 ```bash
-python scripts/infer.py --model_path z-lab/Alpamayo-R1-10B
+python scripts/infer.py --model-path z-lab/Alpamayo-R1-10B
 ```
 
 ## Performance
 
-FlashDrive delivers a comparable speedup on Alpamayo 1 (R1); see the [FlashDrive repository](https://github.com/z-lab/flashdrive) for benchmarks.
+On a single RTX PRO 6000, FlashDrive runs Alpamayo 1 (R1) at **3.9× lower latency** (824 → 211 ms per window) while minADE improves from 1.913 to 1.850. See the [repository](https://github.com/z-lab/flashdrive#performance) for the full benchmark.
 
 ## License
 
-These weights are derived from NVIDIA's Alpamayo release and remain governed by its
-[non-commercial license](https://huggingface.co/nvidia/Alpamayo-R1-10B/blob/main/LICENSE).
-The [FlashDrive](https://github.com/z-lab/flashdrive) inference code is MIT.
+The Alpamayo weights in this repository are released by NVIDIA under the [NVIDIA License](https://huggingface.co/nvidia/Alpamayo-1.5-10B/blob/main/LICENSE), which permits **non-commercial use only** and extends to derivative works. The [FlashDrive](https://github.com/z-lab/flashdrive) inference code is separately released under the [MIT License](https://github.com/z-lab/flashdrive/blob/main/LICENSE).
 
 ## Citation
 
 ```bibtex
 @article{li2026flashdrive,
-      title={FlashDrive: Flash Vision-Language-Action Inference For Autonomous Driving},
-      author={Zekai Li, Yihao Liang, Hongfei Zhang, Jian Chen, Yesheng Liang, Zhijian Liu},
-      year={2026}
+  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}
 }
 ```