CAM++ for AI Pod
Thor R39.2.0 / JetPack 7.2 migration. Model and runtime are separate artifacts.
Original ONNX model is retained unchanged. TensorRT FP32 plan is derived from that model using TensorRT 10.16.2 on Thor, with TF32 disabled. No accuracy beyond the documented regression tests is claimed.
Installed release acceptance passed on the declared Thor target. See reports/performance-report.html and release-summary.json for evidence and limitations.
Release 0.1.1 rejects the earlier FP16 candidate: full-video regression changed 3 speakers to 5 and two embeddings failed cosine >= 0.9999. FP32 restores all 123 intervals and 3 speakers. Runtime archive 0.1.0 is reused unchanged; the selected model plan changes.
Final report
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support