Editing Openai/691c33ba-8898-800c-b30f-1383bae461b1 (section)

==== 4. Very concrete checklist you can try next ====

If I had your codebase in front of me, I’d do this in roughly this order:
# ✅ Confirm all models are .eval() and on GPU, with torch.inference_mode or no_grad around only their forwards.
# ✅ Set: - model.half() + FP16 inputs (or at least clean autocast scopes). - channels_last and pinned memory.
# ✅ Max out useful batch size across: - SAHI tiles - Multiple frames - Pose crops
# ✅ Turn on torch.backends.cudnn.benchmark and, if on torch 2, try torch.compile on RFDetr & RTMPose.
# ✅ Reduce detection frequency using a simple tracker between keyframes.
# ✅ Only compute pose on: - keyframes - key entities - and align with DINO clip sampling.
# ✅ If still too slow → prototype TensorRT/ONNX for one model (probably RFDetr) and measure.

If you want, you can paste a simplified version of your detection+pose loop (just the control flow, no proprietary bits), and I can annotate it with concrete “put autocast here”, “batch these tensors here”, “move this to another process” comments.