Edge AI & Deployment

MobileCLIP2 optimization and validation on Qualcomm QNN for LPCVC 2026.

Model export, graph rewrites, precision tradeoffs, and on-device validation for inference beyond the GPU.

MobileCLIP2 Optimization for Qualcomm QNN LPCVC 2026: deployed MobileCLIP2-S2 as FP16 QNN DLCs on Qualcomm AI Hub’s S22 Family target. Image/text P90 latencies were 14.51/3.99 ms (18.50 ms sum versus 35 ms), with 794/794 profiled operations on NPU.

Export & Graph Optimization

Built a static PyTorch → ONNX opset-16 → QNN pipeline with FastViT reparameterization and normalization, SDPA, and EOS rewrites. Correcting preprocessing raised fractional R@10 from 0.043 to 0.897.

Accuracy & Device Validation

Validated all 56 image and 211 text device outputs at unchanged R@10 of 0.897470, with image/text cosine similarity of 0.999863/0.999998.

Rejected calibrated W8A16 QDQ after image cosine similarity dropped to 0.811, and isolated an S22/v69 DLC-composition failure. These checks informed the deployment choice alongside latency measurements.