vad-trt-edge — deployment checkpoints

Weights accompanying the vad-trt-edge deployment pipeline: TensorRT deployment of VAD (Bench2Drive port) on x86_64 (RTX 4090 Laptop) and NVIDIA Jetson AGX Orin 64GB. Code: the vad-trt-edge GitHub repository (see pipeline docs there for reproduction; the accuracy gates below are defined there).

The upstream base checkpoint is not rehosted here — get vad_b2d_base.pth from the official release: rethinklab/Bench2DriveZoo.

Files

File	Size	SHA-256	What it is
vad_b2d_base_qat_best.pth	225 MB	3288105bf252fcaaa44e09a8e12e17a155100fb31ad60a1c6b29dc6d82e55b3f	VAD-Base (Bench2Drive port) after 2-epoch INT8-aware (QAT) fine-tuning; recovers post-quantization planning accuracy to better than the FP32 reference (L2(1s) 0.6727 vs 0.7300 FP32 / 0.7805 PTQ INT8)
VAD_small_b2d_fast.pth	455 MB	385be4c66d8ed264560b27ffd721f0ecd639e34e71bba8f6df895437c8616683	VAD-Small (custom BEV 100×100 head reduction) trained for 2 epochs as a fast efficiency-validation run; structure-real-time on Orin (~61 ms E2E at INT8, 50 W); not a competitive closed-loop model

Notes

• Checkpoint sizes include FP32 weights + AdamW optimizer state.
• Both files flow through the same export → TRT build → eval chain as the base checkpoint (head I/O contract unchanged).
• Accuracy numbers are the repo's documented eval-set gates (173-frame Bench2Drive mini-set eval split), not closed-loop driving scores.
• VAD-Small is a structure probe, deliberately under-trained; see the repo's docs/model_variants.md for the honesty note and the Base/Small/Tiny comparison.